SOE-YOLO: towards efficient and accurate small object detection in optical remote sensing imagery

1 Introduction

Remote sensing images, captured by satellites, aircraft, or other sensor platforms, provide critical information about the Earth’s surface with extensive scientific and practical applications. This technology plays a crucial role in diverse domains, including urban planning, infrastructure management, national defense, security surveillance, and transportation monitoring. Remote Sensing Object Detection is an intersection of computer vision and remote sensing technologies, aiming to automatically identify and locate specific targets (such as buildings, vehicles, airplanes, ships, and road networks) within remotely sensed images. The process generates both categorical classifications and spatial coordinates, typically represented by bounding boxes. As a core technique in remote sensing image interpretation, its fundamental objective is the efficient and accurate extraction of target objects from complex, large-scale geographical scenes. It is important to note that remote sensing encompasses diverse imaging modalities—including optical, SAR, and infrared imagery—each presenting distinct challenges for target detection Zhang et al. (2026d), Zhang et al. (2026c). While this study specifically focuses on optical remote sensing imagery, the broader field also addresses speckle-robust and quantity-aware detection in SAR imagery (Zhang et al., 2026b; Zhang et al., 2026a), which represent fundamentally different detection paradigms.

In recent years, artificial intelligence image processing technology represented by Deep Learning (Wang et al., 2021; Khan et al., 2021) has made significant progress, which has driven the rapid development of remote sensing target detection algorithms based on a convolutional neural network (CNN). This breakthrough has overcome the accuracy and speed bottlenecks inherent in traditional manual feature extraction methods. Through end-to-end feature learning mechanisms, deep learning automatically extracts multilevel, high-semantic abstract features. It demonstrates enhanced capability in accurately capturing the texture, shape, and contextual information of targets in complex scenarios. Moreover, deep learning models (such as CNN and Transformer (Vaswani et al., 2017)) are naturally equipped with the ability to deal with multi-scale and multidirectional targets. By implementing structures such as feature pyramids (Tan et al., 2020) and attention mechanisms (Ouyang et al., 2023; Hu et al., 2018), they effectively address challenges in small target detection and rotational target localization while exhibiting stronger robustness against complex background interference (such as cloud cover, shadows, and spectrally similar objects). When integrating with data augmentation techniques (Ouyang et al., 2023) and adaptive fusion strategies (including multimodal data fusion) (Lu et al., 2019), these models achieve substantial improvements in detection accuracy and generalization capability. These advantages have established deep learning as the core technology for remote sensing target detection, which enables the transition from traditional rule-driven approaches to data-driven paradigms and is widely used in rescue at sea, reconnaissance systems, and urban planning. Furthermore, the rapid advancement of satellite technology has significantly enhanced the quality and resolution of remote sensing imagery, especially for panchromatic satellite images (PAN images) (Gao et al., 2022), which usually have a high spatial resolution, making research on deep learning-based target detection algorithms for PSI of significant application value.

Existing researches (Xie et al., 2023; Qu et al., 2025; Betti and Tucci, 2023) have significantly improved object detection performance in remote sensing imagery through diverse strategies, including multi-level feature fusion and attention mechanisms. Recent advances in SAR small-target detection have introduced innovative paradigms worth noting: instance segmentation-based approaches for small-scale ship targets (Zhang Y. et al., 2026), attention-matrix despeckling with Haar wavelet reconstruction (Zhang et al., 2025), and adaptive fast refocusing techniques for complex-motion targets (Xu et al., 2025b). While these methods operate in the SAR domain, their underlying strategies—progressive feature aggregation and attention-guided reconstruction—exemplify design principles that are transferable across modalities. However, when confronting the extreme scale variations, arbitrary orientations, and complex background clutter inherent in aerial and satellite imagery, current mainstream approaches still exhibit several critical bottlenecks that require urgent resolution:

• Spatial Misalignment and Detail Loss during Downsampling: Mainstream network architectures heavily rely on standard symmetric downsampling operations (e.g., strided convolutions or pooling) to compress feature map resolution and enlarge receptive fields. However, these operations inherently induce sub-pixel spatial feature misalignment and truncate edge information. For minute targets occupying merely a few pixels, this spatial displacement accumulates progressively across deeper layers, causing an irreversible loss of fine-grained spatial cues and making precise localization exceptionally difficult.
• Degradation of High-Frequency Signals: Deep spatial-domain convolutional neural networks inherently act as low-pass filters during continuous feature extraction. This mechanism inadvertently smooths out the crucial high-frequency details of small targets, such as sharp structural boundaries and local contrast variations. Consequently, the extremely weak signals of tiny targets (e.g., distant vehicles or scattered ships) are easily submerged by overwhelming optical backgrounds sharing similar textures, such as urban terrain, vegetation, or cloud shadows.
• Suboptimal Fusion for Multi-Oriented Targets: Remote sensing targets frequently appear in arbitrary orientations and drastic aspect ratios (e.g., long bridges or rotated aircraft). Traditional multi-scale fusion structures often lack dimension-aware dynamic adaptation mechanisms. They struggle to adaptively allocate attention weights across different spatial and channel dimensions, hindering the model’s ability to discriminate distinctively oriented targets against highly complex and chaotic backgrounds.
• Semantic Gaps in Feature Concatenation: Existing multi-scale feature pyramids (e.g., FPN, PANet) typically employ primitive concatenation to fuse cross-level features. This brute-force merging creates a severe semantic gap: shallow feature maps contain vital fine-grained spatial details but are heavily corrupted by background noise, whereas deep features possess strong global semantics but lack precise spatial boundaries. Direct concatenation frequently causes the fragile feature responses of tiny targets to be completely drowned out by the expansive, cluttered background noise propagated from the shallow layers.

To systematically address these interrelated bottlenecks and strike a state-of-the-art balance between detection robustness and computational efficiency, this study proposes SOE-YOLO (Small Object Enhanced-YOLO11), a novel and lightweight object detection framework. The main contributions of this study are structured as follows:

• We propose an Asymmetric Padding Downsampling (APDS) module that processes input features through four branches with distinct padding directions (top-left, bottom-right, bottom-left, top-right). This design enables the convolution kernel to focus on contextual information from different edge regions. This padding offset indirectly expands the effective receptive field of the convolutional kernel. By integrating channel attention mechanisms with explicit spatial position modeling, the integrated design employs a synergistic mechanism of regional cross-attention, position-aware deconvolution, and feature fusion. This approach achieves the core subsampling function (reducing resolution while expanding the receptive field) while actively and structurally counteracting feature displacement effects. This significantly enhances the multi-scale feature fusion capability of subsequent deep networks, improving their robustness while systematically mitigating information loss and feature distribution bias caused by traditional downsampling operations.
• We propose a novel High-Frequency Contrast Enhancement (HFCE) module tailored for optical remote sensing imagery. The HFCE module is seamlessly integrated into the YOLO bottleneck architecture—Cross-Stage Partial with two convolutions (C3k2), which is the default bottleneck block in YOLO11 featuring two lightweight convolutional branches for efficient feature extraction—(forming the C3k2-HFCE block). It leverages fixed-weight physical priors, hierarchical multiscale fusion, and dual-attention mechanisms to explicitly preserve high-frequency features of small targets while ensuring extreme computational efficiency.
• We introduce the Dimension-Aware Selective Integration Module (DASI) (Xu et al., 2024), which significantly improves the detection performance in complex scenes (especially for small targets and multi-orientation targets) through dynamic fusion of multi-scale features and adaptive weight allocation. Through multi-scale patch fusion architecture and dynamic gating mechanisms, this module effectively addresses challenges in remote sensing target detection such as missed detection of small targets, multi-orientation adaptation, and interference from complex backgrounds.
• We propose the Contextual Small-object Attention Fusion (CSAF) module to replace the primitive concatenation in the network neck. The CSAF module bridges the semantic gap through cross-level channel correlation and strongly suppresses clustered backgrounds via a competitive spatial attention mechanism.

The four proposed modules should not be viewed as isolated improvements but as an integrated pipeline addressing complementary stages of the small object detection problem. From a systems perspective, the framework follows a “Preserve–Enhance–Fuse–Select” (PEFS) design philosophy: Preserve (APDS): At the earliest stage of feature extraction, the backbone’s downsampling operations risk destroying the spatial integrity of small targets before any subsequent module can process them. APDS is placed at the transition between backbone stages to proactively preserve spatial positional information during resolution reduction, ensuring that downstream modules receive structurally intact feature maps. Enhance (C3k2-HFCE): Even after spatial preservation, the inherently weak high-frequency signatures of small targets require explicit amplification to compete with dominant low-frequency background textures. C3k2-HFCE operates within the backbone’s bottleneck blocks to enhance the signal-to-noise ratio of small-target features at multiple receptive field scales, making tiny objects more distinguishable before they enter the neck. Fuse (DASI): Once features are preserved and enhanced, multi-scale fusion must reconcile the tension between high-resolution detail and high-level semantics. DASI operates at the neck junctions where feature maps from different backbone levels meet, dynamically weighting the contribution of low-level (detail-rich) and high-level (semantic-rich) features on a per-channel basis. This ensures that small-target details are not overridden by coarse semantic features during multi-scale aggregation. Select (CSAF): The final fused feature maps, although enriched with multi-scale information, still contain substantial background clutter. CSAF acts as a gating mechanism at the neck’s output, employing cross-level semantic recalibration and competitive spatial attention to selectively amplify target-relevant features while suppressing background noise, before the features are passed to the detection head.

This PEFS pipeline is inherently serial and mutually reinforcing: without APDS’s spatial preservation, HFCE’s enhancement would operate on misaligned features; without HFCE’s enhancement, DASI’s fusion would fuse indistinguishable weak signals; without DASI’s adaptive fusion, CSAF’s attention would compete against semantically impoverished features. The empirical ablation results (Table 3) confirm this synergistic relationship: the combined improvement of APDS + C3k2-HFCE (+4.3% mAP50, Experiment 6) exceeds the sum of their individual gains (+2.3% and +2.0%, Experiments 2–3), and the full PEFS pipeline (+7.4%, Experiment 8) yields a gain that exceeds the linear sum of individual contributions (+7.1%). This super-additive effect constitutes quantitative evidence that the four modules address distinct, complementary failure modes in the detection pipeline, validating the unified theoretical motivation.

2 Related works

The evolution of object detection in remote sensing imagery has been significantly driven by adaptations of the YOLO framework. To address domain-specific challenges such as large scale variations, small object sizes, and complex backgrounds, researchers have introduced a range of structural innovations. These improvements can be broadly categorized into four technical directions: (1) detection head restructuring and pyramid scaling, (2) feature enhancement and fusion strategies, (3) attention mechanisms and context modeling, and (4) lightweight design and efficiency optimization.

2.1 Detection head restructuring and pyramid scaling

A critical line of research focuses on adapting the feature pyramid structure and detection head configuration to better align with the scale distribution of targets in remote sensing images. The standard P3-P5 head design in general-purpose detectors often leads to spatial information loss for small objects, as targets can become sub-pixel in deeper layers with large strides (e.g., stride 32 in P5).

The CMS-YOLOv7 (Qin et al., 2024) model addresses this by adding a high-resolution P2 detection layer to enhance small object detection while removing the P5 layer to reduce computational redundancy from excessive downsampling. This is coupled with an SPD-Conv module that uses space-to-depth transformations to preserve spatial details of objects smaller than 32 pixels.

2.2 Feature enhancement and fusion strategies

Enhancing the discriminative power of features, especially for small and blurry objects, is another prominent research direction. This involves designing modules to capture richer contextual information and improving the fusion of multi-scale features.

ADS-YOLO (Gui and Geng, 2024) incorporates a DWR module into the C2f backbone, where dilated convolutions in bottleneck structures expand the receptive field without increasing parameters significantly, thus capturing multi-scale contextual features more effectively.

Feature fusion mechanisms are crucial for combining information from different levels. YGNet (Song and Gao, 2025) proposes a dual-path architecture with HLS-PAN for hierarchical feature interaction, enabling bidirectional fusion across scales through selective channel attention. Similarly, FE-YOLO (Wang et al., 2023) redesigns the neck with an FFD module that aggregates multi-scale global features from backbone layers and redistributes them to prediction heads, mitigating information loss in non-adjacent layer fusion and enhancing fine-grained detail retention for small targets.

2.3 Attention mechanisms and context modeling

To mitigate interference from complex backgrounds in remote sensing images, attention mechanisms have been widely integrated to help models focus on salient regions.

AG-YOLO (Wang et al., 2025) incorporates an attention-guided branch in the detection head that generates saliency maps from shallow features, dynamically weighting feature maps during aggregation to suppress background noise. It also extends the detector by adding a rotation angle parameter for better orientation adaptation.

SED-YOLO (Wei et al., 2025) employs an EMA (Ouyang et al., 2023) module, which replaces standard convolutions with switchable atrous convolutions for adaptive receptive field adjustment, and a DyHead module that dynamically weights multi-scale features based on spatial-channel relationships in the detection head.

Beyond the optical domain, innovative context modeling paradigms have also emerged in SAR remote sensing. The swarm learning framework (Zhang et al., 2026c) proposes a perception–retrieval–localization pipeline for ship detection, leveraging target-dependent information across scenes to enhance detection performance. The glance-focus-gaze paradigm (Zhang et al., 2026a) introduces a panorama-population-individual progressive screening strategy inspired by eagle-eye vision, achieving efficient multi-stage target capture. Recently, Zhang et al. proposed a divergence-to-concentration and population-to-individual progressive approaching paradigm for SAR ship detection (Zhang et al., 2026e), further advancing the progressive screening strategy. A triple-level sparsity awareness framework (Zhang et al., 2026f) simultaneously considers imaging sparsity and target scattering sparsity for robust SAR marine surveillance. Comprehensive surveys on deep learning-based ship detection in SAR imagery (Xu et al., 2026) and diffusion-model-based SAR ship instance segmentation (Xu et al., 2025c) have also systematized the field, providing valuable cross-modal insights. These paradigms demonstrate the value of structured context modeling and progressive screening strategies that could potentially inform optical remote sensing detection as well.

While these attention mechanisms are effective, challenges remain when target features are highly similar to the background (e.g., vehicles in vegetation), where traditional attention may struggle to differentiate effectively, potentially leading to decreased accuracy.

2.4 Lightweight design and efficiency optimization

For real-time applications and deployment on resource-constrained platforms, model lightweighting and efficiency optimization are essential.

YOLO-S (Betti and Tucci, 2023) achieves edge-device compatibility through a reshape-passthrough layer and skip connections that reuse low-level features, resulting in a model with 90% fewer parameters than Tiny-YOLOv3 while maintaining higher accuracy.

YOLO-RS (Xie et al., 2023) optimizes YOLOv4 with a Lightnet backbone based on depthwise separable convolutions and an ASFF (Qiu et al., 2022) module that assigns learnable weights to multi-scale features during fusion, achieving a favorable balance between speed and accuracy.

2.5 Summary

In summary, existing research has significantly improved object detection performance in remote sensing images—particularly for small targets and complex backgrounds—through diverse strategies such as multi-level feature fusion, attention mechanisms, and specialized module designs. However, these approaches still face several inherent challenges. For instance, multi-scale feature fusion architectures (e.g., FPN, PANet) heavily rely on standard downsampling operations (e.g., stride convolution or pooling). These operations reduce feature map resolution while often introducing feature misalignment, which is particularly detrimental to the precise localization of minute objects. Furthermore, while dynamic attention mechanisms (e.g., CBAM (Woo et al., 2018), coordinate attention) effectively enhance feature representation capabilities, their ability to model spatial structural context remains limited when handling targets with extreme aspect ratios or arbitrary orientations (e.g., bridges, ships). At the label assignment and loss function level, the inherent high sensitivity of traditional Intersection over Union (IoU) and its variants on minute objects also limits further improvements in detection performance.

3 Materials and methods

3.1 SOE-YOLO overview

In order to solve the problems of small object detection, multiorientation object detection, and complex background interference in remote sensing image target detection, this paper proposes an object detection network based on an improved YOLO11n model. The baseline YOLO11n architecture comprises several key components: the C3k2 (Cross-Stage Partial with two convolutions) bottleneck blocks for efficient feature extraction, the SPPF (Spatial Pyramid Pooling—Fast) module for multi-scale context aggregation via successive max-pooling, and the C2PSA (Cross-Stage Partial with two convolutions and Position-Sensitive Attention) module that incorporates spatial and channel attention to enhance positional feature representation. In remote sensing image object detection tasks, small objects are distributed in multiple orientations. Algorithm design necessitates a careful balance between detection accuracy and computational speed. To improve detection accuracy without reducing detection speed, this research further improves the one-stage object detection algorithm YOLO11.

In the backbone network, we propose an asymmetric padding downsampling module that forces the convolutional kernel to focus on the contextual information from different edge regions through four different padding directions (upper-left, lower-right, lower-left, and upper-right), which are processed by each of the four branches. The offset of this padding indirectly expands the effective receptive field of convolution kernels, capturing contextual associations over a wider range. Each branch independently extracts features from different directions and retains multi-angle feature diversity after concentration. This downsampling module employs the concept of asymmetric padding and multi-branch fusion to preserve the multi-directional features, adapt to multi-directional objects, reduce information loss, and solve the problems of multi-directional and small-sized objects in remote sensing scenarios. Meanwhile, the Dimension-Aware Selective Integration Module (DASI) is introduced, which effectively solves the problems of small target leakage, multi-direction adaptation and complex background interference in remote sensing object detection through multi-scale feature dynamic fusion with adaptive weight allocation, multi-scale block fusion, and dynamic gating mechanisms.

In the neck network, we propose the High-Frequency Contrast Enhancement (HFCE) module. In standard convolutional neural networks, consecutive down-sampling operations inherently act as low-pass filters, inadvertently smoothing out the crucial high-frequency details (e.g., edges and local contrast) of tiny remote sensing targets. Inspired by the physical priors used in infrared target detection, the HFCE module introduces a zero-parameter, fixed-weight Center-Surround contrast mechanism as a strong inductive bias. By integrating this parameter-free high-pass filter with hierarchical multiscale fusion and dual-attention refinement, the HFCE module explicitly forces the backbone to highlight blob-like small targets and sharp edges. This design exceptionally enhances the network’s perception of variably sized objects against complex optical backgrounds without incurring additional computational overhead. Furthermore, to overcome the severe semantic misalignment and background interference during multi-scale feature fusion, we design the Contextual Small-object Attention Fusion (CSAF) module to replace the primitive concatenation in the network neck. Conventional concatenation directly merges shallow spatial details with deep semantic features, which often causes the weak signals of tiny targets (occupying merely a few pixels) to be drowned out by overwhelming background noise. The proposed CSAF bridges this semantic gap by establishing a cross-level channel correlation matrix for bidirectional semantic recalibration. Crucially, it incorporates a competitive spatial Softmax attention mechanism that enforces a strict probability distribution across the spatial dimension. This ruthless competition guarantees that the network allocates the scarcest attention weights exclusively to salient small targets while actively suppressing expansive cluttered backgrounds, fortified by a residual compensation path to prevent target feature diminishment. The general structure of the network is shown in Figure 1.

3.2 Asymmetric padding downsampling module

Feature shift originates from pixel position deviations in feature maps caused by stride configurations and padding strategies during downsampling. Edge features within the feature maps may be incorrectly allocated, leading to a decline in detection accuracy. As illustrated in Figure 2, traditional symmetric padding convolutions truncate edge features and induce a 0.5-pixel offset in higher-resolution inputs (with padding indicated by the blue dashed region), disrupting the positional consistency of small object features (center offset of the receptive field).

Consider a standard YOLO11n backbone that performs N = 5 stride-2 downsampling operations (through convolutional layers with stride two and MaxPool layers), progressively reducing a 320 × 320 input to 10 × 10 at the deepest stage. At each downsampling step k ∈ { 1,2 , … , 5 }, the spatial resolution is halved: H k = H k − 1 / 2, W k = W k − 1 / 2, where H 0 = W 0 = 320.

Let p 0 = ( x 0 , y 0 ) ⊤ ∈ R 2 denote the continuous coordinate of a small target center in the input image. Under an ideal alias-free downsampling operation, the coordinate at layer k is p k = p k − 1 / 2. However, in practice, each stride-2 convolution with kernel size three and padding one introduces a sub-pixel alignment error δ k = ( δ k , x , δ k , y ) ⊤ resulting from two factors:

When the kernel size K = 3 with stride S = 2 operates at the spatial boundary, the standard symmetric padding of one pixel on each side creates an imbalance: the receptive field center at odd-indexed positions (e.g., x = 1) is offset by 0.5 pixels relative to the grid-aligned position. Formally, the output coordinate at position i after a convolution with padding P, kernel size K, and stride S is given by Equation 1:For K = 3, S = 2, P = 1, this yields x out ( i ) = 2 i + 1 − 1 = 2 i, which is grid-aligned. However, when the input dimension is odd (as occurs at intermediate resolutions), the floor operation in the convolution implementation introduces a 0.5-pixel offset at the boundary.

x out i = S ⋅ i + K 2 − P (1)

Each downsampling stage k incurs a displacement error ϵ k that is propagated and amplified through subsequent stages. The effective displacement at layer k, when mapped back to the original input resolution, is:where the factor 2 j − 1 accounts for the resolution ratio between layer j and the input.

Δ p k input = ∑ j = 1 k 2 j − 1 ⋅ ϵ j (2)

For a typical 5-layer YOLO11n backbone with K = 3, S = 2, and P = 1 at each stage, the per-layer displacement error can be modeled as ϵ j ≈ ( − 0.5 , − 0.5 ) pixels at the corresponding feature map resolution. Equation 3 defines the output, Substituting into Equation 2:

Δ p 5 input = ∑ j = 1 5 2 j − 1 ⋅ − 0.5 , − 0.5 = 1 + 2 + 4 + 8 + 16 ⋅ − 0.5 , − 0.5 = 31 ⋅ − 0.5 , − 0.5 = − 15.5 , − 15.5 pixels (3)

The total Euclidean displacement is ‖ Δ p 5 input ‖ 2 = 15 . 5 2 + 15 . 5 2 ≈ 21.9 pixels. For a representative small target of size 16 × 16 pixels, this represents a positional drift exceeding the target’s own dimensions. When additionally accounting for the accumulated errors from MaxPool layers and nonlinear activation-induced deformations, the effective cumulative displacement reaches approximately 31.5 pixels in total spatial offset (combining both x and y components). This mathematically explains why small objects—whose feature representations occupy only a few pixels at deep layers—are particularly vulnerable to even sub-pixel alignment errors: a 0.5-pixel error at layer 5, when mapped back, corresponds to a 16.0-pixel displacement in the input domain, which can entirely shift the receptive field center away from the target.

Remote sensing images exhibit high resolution with dense small objects (e.g., vehicles, vessels). Conventional downsampling methods like stride-2 convolution severely degrade edge-positioned targets due to information loss from symmetric padding (e.g., padding = 1).

In classical downsampling operations such as stride convolution or pooling, the limited receptive field and monolithic aggregation of local information often cause feature shifts in both spatial and channel dimensions. This disrupts the stability of multi-scale feature fusion and gradient propagation. This paper proposes an asymmetric padding downsampling module, whose architecture is shown in Figure 3. By employing four independent branches and dynamic adaptive spatial attention, it achieves directional asymmetric downsampling. While preserving local structural information, this module introduces a spatio-channel collaborative attention mechanism to adaptively calibrate feature distribution.

The mechanism of the Asymmetric Padding Downsampling Module (APDS) is as follows: First, the input features undergo parallel branches of max and average pooling to simultaneously capture salient responses and global context within the features. In this branch, the input features are subjected to max and average pooling respectively, followed by weighted fusion. Next, 1 × 1 convolutions generate channel attention, where the kernel size is adaptively determined by a nonlinear mapping of the channel dimension. The resulting channel weights are then multiplied with the input features. These features are subsequently fed into four position-aware asymmetric padding convolutions (PADConv) branches, dividing the features into four independent subregions (top-left, top-right, bottom-left, bottom-right). Each undergoes separate convolution processing to explicitly model feature patterns across different spatial locations, mitigating local structural blurring and positional information loss caused by downsampling. The four position-aware asymmetric padding convolutions (PADConv) employ a stride of two and a 3 × 3 convolution kernel. PADConv1 pads the top-left region, preserving the bottom-right feature; PADConv2 pads the bottom-right region, preserving the top-left feature; PADConv3 pads the bottom-left region, preserving the bottom-right feature; PADConv4 pads the top-right region, preserving the top-left feature. After passing through PADConv, features from each subregion undergo channel concatenation to facilitate information exchange and integration. This enables the module to dynamically calibrate the contribution weights of features from different positions. Finally, batch normalization and the SiLU activation function further stabilize the distribution of output features, enhancing nonlinear expressive capabilities. This architecture effectively suppresses performance degradation caused by feature shifts, improving the model’s robustness across multi-scale tasks.

The core innovation of this downsampling module lies in integrating channel attention mechanisms with explicit spatial position modeling. Through a combined approach of regional cross-attention, position-aware deconvolution, and feature fusion. This integrated design achieves the core function of downsampling (reducing resolution and increasing the receptive field) while actively and structurally counteracting feature shifts. Consequently, it enhances the subsequent deep network’s ability to fuse multi-scale features and improves its robustness, systematically mitigating the information loss and feature distribution bias caused by traditional downsampling operations.

The error-correction process is visually demonstrated in Figure 4, where directional offsets are systematically compensated by the algorithm. As analytically illustrated in Figure 4, the mechanism of APDS can be summarized as follows. In the vertical-direction analysis, asymmetric padding is applied separately to the left side (Figure 4a) and the right side (Figure 4b). A stride-2 convolution (kernel size = 3) is applied to generate directionally specialized features. To mitigate misalignment, channel-wise concatenation is employed to produce offset-corrected feature vectors (Figure 4c). For reference, the original input vector (Figure 4d) exhibits symmetric kernel coverage, where blue and pink blocks indicate regions convolved twice and three times, respectively. The theoretical innovation lies in the strategic repositioning of receptive fields via asymmetric padding, which enables branch-specific feature capture from divergent spatial locations. Through cross-branch complementary fusion, directional shifts are effectively neutralized, thereby mathematically restoring symmetry in the aggregated output.

Assuming an input feature map X ∈ R C × H × W, where C denotes the number of channels, H the height, and W the width, we set C = 1, H = 16, W = 16. The input X is processed using both a standard stride-2 convolution with kernel size three and the proposed APDS method, in which kernel activation values are set to 0.25 and target activation values to 1. In this work, feature map positions are represented using matrix indices ( i , j ), where (0,0) corresponds to the top-left corner.

Figures 5, 6 present the output heatmaps after processing by APDS when the target is located at position (1,1) and the central region (7:9,7:9), respectively. The results demonstrate that APDS effectively alleviates feature misalignment compared to ordinary convolution. Moreover, the asymmetric padding strategy preserves edge information across different directions, reducing information loss during downsampling—a particularly valuable property for detecting small objects.

3.3 High-Frequency Contrast Enhancement (HFCE) module

Standard spatial-domain convolutional neural networks (CNNs) tend to act as low-pass filters during the feature extraction process, which inadvertently smooths out high-frequency details and hinders the ability to detect small, variably sized, and low-contrast targets against complex backgrounds. In optical remote sensing datasets such as VEDAI, RSOD, and NWPU VHR-10, small targets (e.g., tiny vehicles or scattered ships) often occupy only a few pixels and are easily submerged by overwhelming background noise.

To address this challenge, inspired by the multiscale atrous contrast convolution originally designed for infrared target detection (Xu M. et al., 2025), we propose a novel High-Frequency Contrast Enhancement (HFCE) module tailored for optical remote sensing imagery, the structure of the module is shown in Figure 7. The HFCE module is seamlessly integrated into the YOLO bottleneck architecture (forming the C3k2-HFCE block). It leverages fixed-weight physical priors, hierarchical multiscale fusion, and dual-attention mechanisms to explicitly preserve high-frequency features of small targets while ensuring extreme computational efficiency.

3.3.1 Zero-parameter fixed-weight priors for high-frequency extraction

Unlike conventional convolutions that learn weights from scratch, the HFCE module introduces handcrafted Center-Surround contrast filters as a strong inductive bias. These filters mimic the Difference of Gaussians (DoG) operator, where the central region acts as an excitatory field with positive weights, and the surrounding region acts as an inhibitory field with negative weights.

To prevent overfitting on smaller datasets and strictly control the model’s parameter complexity, the weights of these contrast filters are frozen during the training phase (i.e., gradients are not updated). Let k be the kernel size, the weights are defined such that the sum of the entire kernel is zero, as formulated in Equation 4:

W center = 1 A center , W surround = − 1 k 2 − A center (4)

where A center represents the area of the central positive region. This zero-parameter design acts as an explicit high-pass filter, forcing the network to highlight blob-like small targets and sharp edges without incurring additional learning parameters.

The fixed-weight design is grounded in signal processing theory. The Laplacian-of-Gaussian (LoG) operator and its approximation via Difference of Gaussians (DoG) are provably optimal for detecting blob-like structures at a given scale σ under the scale-space framework (Lindeberg, 1994). Specifically, the LoG response for a circular blob of radius r is maximized when σ = r / 2, and the zero-sum constraint ( ∑ W = 0 ) ensures DC component removal, making the filter invariant to illumination shifts that are prevalent in remote sensing imagery (e.g., varying sun angles, atmospheric scattering). A learnable k × k convolution, initialized randomly, must discover this structure from data; however, on small-scale remote sensing datasets (e.g., VEDAI with only 1,210 images), the optimization landscape is sufficiently noisy that learned filters often converge to suboptimal local minima that mix low-frequency and high-frequency responses, degrading small-target selectivity. Furthermore, the fixed-weight prior reduces the number of trainable parameters by k 2 per filter, which is non-trivial given the multiple parallel branches in HFCE.

To empirically validate this design choice, we conducted an ablation experiment comparing the fixed-weight HFCE against a learnable-weight variant (denoted HFCE-Learnable), where the contrast filter weights are initialized with the same DoG-inspired values but allowed to update via backpropagation. As shown in Table 1, on the VEDAI dataset, the fixed-weight HFCE achieves an mAP50 of 62.4% and mAP50:95 of 37.2%, while the learnable variant yields 61.9% and 36.8%, respectively. The degradation of 0.5% mAP50% and 0.4% mAP50:95 with the learnable variant confirms that: (1) the handcrafted priors already encode near-optimal high-frequency extraction kernels for small target enhancement, and (2) allowing gradient updates introduces additional degrees of freedom that are prone to overfitting on limited training data, thereby diluting the targeted high-pass selectivity. This empirical result is consistent with the theoretical prediction that, when the structural prior is well-defined and the training data is limited, hard-coding the prior outperforms learning it from scratch.

TABLE 1. Ablation study comparing fixed-weight and learnable-weight HFCE variants on the VEDAI dataset.

Bold values indicate the best performance or optimal results for each evaluation metric across the compared models or configurations (e.g., the highest mean Average Precision [mAP] and Frames Per Second [FPS], or the fewest parameters [Params] and computation costs [GFLOPs]).

3.3.2 Hierarchical multiscale fusion

Remote sensing targets exhibit drastic intra-class scale variations. To capture a wide range of contextual information, the input feature map X is first projected via a 1 × 1 convolution and then evenly split into S groups (we set S = 4) along the channel dimension, denoted as X 0 , X 1 , X 2 , X 3. We construct a Res2Net-style hierarchical progressive fusion which is expressed by Equation 5:

Y 0 = X 0 Y 1 = R e L U C o n v 1 X 1 Y 2 = R e L U C o n v 2 X 2 + Y 1 (5)

where C o n v 1 and C o n v 2 represent depthwise convolutions initialized with the aforementioned fixed-weight contrast priors of different scales. This split-and-concatenate strategy inherently provides combinational receptive fields, perfectly matching the variable scales of remote sensing targets.

3.3.3 High-low frequency aggregation and dual-attention refinement

For the final feature group, to further suppress complex urban or terrain backgrounds, we propose a dual-path aggregation combining a low-pass smoothing filter and a dilated high-pass edge-enhancement filter which is expressed by Equation 6:

Y 3 = R e L U B N C o n v mean X 3 + Y 2 + C o n v edge X 3 + Y 2 (6)

After concatenating Y 0 through Y 3 and applying a 1 × 1 fusion convolution, the fused feature inevitably contains residual background noise. Therefore, we sequentially append a Channel Attention (CA) and a Spatial Attention (SA) mechanism to perform dynamic feature recalibration. The CA selectively emphasizes channels sensitive to high-frequency target features, while the SA highlights spatial regions containing potential small targets. The final output is obtained through a residual connection with the original input X, ensuring stable gradient propagation.

3.4 Dimension-aware selective integration module

During multiple downsampling processes in remote sensing object detection, high-dimensional features may lose information about small objects, while low-dimensional features may not provide sufficient contextual information. To address this issue, this paper introduces a novel channel division selection mechanism that enables DASI to adaptively select appropriate features for fusion based on object size and characteristics. DASI dynamically adjusts the fusion ratio of high- and low-dimensional features, selecting the most suitable feature combination according to target size and contextual information. This helps retain small object details while leveraging semantic information from higher-level features. Small object information is easily lost during multiple downsampling operations. DASI supplements detailed information by fusing low-dimensional features (which are potentially from lower layers with higher resolution), preventing small objects from disappearing during the downsampling process, and mitigating information loss. High-dimensional features typically contain more contextual and semantic information but may lack details. Through selective integration, DASI enables the model to utilize global information from high-level features, helping distinguish small objects from complex backgrounds and enhancing contextual awareness. Through channel splitting and parallel processing, the DASI module reduces computational complexity while improving the flexibility and effectiveness of feature fusion, enabling the network to efficiently capture target features at different scales and enhancing feature fusion efficiency, as illustrated in Figure 8.

DASI takes multi-level feature maps as input. Low-level feature maps, from shallow networks, are larger in size and retain high-resolution details (edges, textures, and small target pixels). High-level feature maps, from deep networks, are smaller and contain rich semantics (target shapes, categories). The current-level feature map is the intermediate feature map to be fused.

To reduce computational complexity, DASI evenly splits each feature map into four sub-blocks along the channel dimension. If the total number of channels C is not divisible by 4, the last sub-block is zero-padded. Each sub-block is processed independently. As formulated in Equation 7:

u i ′ = α i ⋅ l i + 1 − α i ⋅ h i i = 1,2,3,4 (7)

where l i, h i, and u i ′ are the corresponding sub-blocks, and α i is the gate control weight for the sub-block. The gate weight α i is dynamically calculated via a cross-dimensional attention mechanism. The current sub-block is concatenated with the corresponding low-level and high-level sub-blocks to form a joint feature C o n c a t ( u i , l i , h i ). If the channel numbers differ, they are aligned via a 1 × 1 convolution. The concatenated features are then compressed via Global Average Pooling (GAP) and α i are generated through a Fully Connected (FC) layer and a Sigmoid activation function, as formulated in Equation 8:

α i = σ F C G A P C o n c a t u i , l i , h i (8)

Here, σ is the Sigmoid function that constrains α i to the [0,1] interval. When α i approaches 1, the current sub-block relies more on low-level features, enhancing detail preservation for small targets or high-texture regions. When α i approaches 0, the sub-block relies more on high-level features, strengthening semantics for complex backgrounds or blurry targets.

The sub-blocks are then fused through the following method, as formulated in Equation 9:

u i ′ = α i ⋅ l i + 1 − α i ⋅ h i (9)

The four sub-blocks { u 1 ′ , u 2 ′ , u 3 ′ , u 4 ′ } are concatenated along the channel dimension to restore the original number of channels. Finally, a 1 × 1 convolution (Conv) + Batch Normalization (BN) + ReLU activation is used for channel interaction and non-linear enhancement, producing the final fused feature F out.

3.5 Contextual Small-object Attention Fusion (CSAF) module

In standard YOLO architectures, including YOLO11, the feature pyramid network (FPN) and path aggregation network (PAN) rely on simple concatenation operations to fuse multi-scale feature maps. However, direct concatenation suffers from severe semantic gaps when dealing with optical remote sensing imagery. Shallow feature maps contain rich high-frequency details but are cluttered with background noise, whereas deep feature maps possess strong semantic information but lack fine-grained spatial resolution. For tiny remote sensing targets (e.g., small vehicles and ships) occupying merely a few pixels, simple concatenation inevitably causes the weak target features to be drowned out by overwhelming background responses.

To overcome this bottleneck, we propose the Contextual Small-object Attention Fusion (CSAF) module to replace the primitive concatenation in the network neck. Its structure is shown in Figure 9. The CSAF module bridges the semantic gap through cross-level channel correlation and strongly suppresses clustered backgrounds via a competitive spatial attention mechanism. Given two input feature maps from different scales, f 1 and f 2, the CSAF module first aligns their channel dimensions to C out using 1 × 1 convolutions. The subsequent fusion process consists of three core stages:

3.5.1 Bridging the semantic gap via cross-level interaction

In conventional feature pyramids, fusing features across different network depths inherently suffers from semantic misalignment. Shallow features ( f 1 ) preserve crucial high-resolution details for small object localization but are heavily corrupted by background noise due to their limited receptive fields. Conversely, deep features ( f 2 ) encode robust global semantics but lose fine-grained spatial boundaries after multiple downsampling stages. To effectively bridge this semantic gap, we propose a cross-level interaction mechanism that dynamically aligns and exchanges information between different scales prior to physical fusion.

Specifically, we first extract the global context of both scales by applying Global Average Pooling (GAP) and Global Max Pooling (GMP) to the flattened f 1 and f 2. These pooled descriptors are processed by a multi-layer perceptron (MLP) to generate robust channel-wise representations a 1 , a 2 ∈ R B × C × 1. To model the inter-dependencies between the deep and shallow features, we compute a cross-channel correlation matrix via matrix multiplication, as formulated in Equation 10:

W cross = a 1 ⋅ a 2 T (10)

where W cross ∈ R B × C × C serves as a dense semantic bridge. It explicitly quantifies how each channel in the shallow feature correlates with every channel in the deep feature. Subsequently, this correlation matrix is normalized along the channel dimension via a Softmax function and multiplied back into the original features to perform bidirectional semantic recalibration, as formulated in Equation 11:

f ̂ 1 = S o f t m a x W cross ⋅ f 1 f ̂ 2 = S o f t m a x W cross T ⋅ f 2 (11)

Through this cross-level interaction, the network achieves a synergistic feature exchange: the shallow feature f ̂ 1 absorbs high-level semantic guidance from f 2 to filter out background clutter, while the deep feature f ̂ 2 retrieves high-frequency spatial cues from f 1 to restore the precise boundaries of tiny targets. This mechanism fundamentally replaces abrupt feature concatenation with a smooth, context-aware semantic fusion.

3.5.2 Competitive spatial attention for background suppression

Remote sensing images are characterized by expansive fields of view and complex interferences. To accurately localize small targets, we design a competitive spatial attention mechanism. For the recalibrated features f ̂ 1 and f ̂ 2, we compute the spatial attention maps by concatenating their channel-wise mean and max values, followed by two consecutive 3 × 3 spatial convolutions. Crucially, instead of using a standard Sigmoid activation, we flatten the spatial dimensions and apply a spatial Softmax function, as formulated in Equation 12:

M i = S o f t m a x C o n v 3 × 3 C o n v 3 × 3 A v g f ̂ i ; M a x f ̂ i , d i m = − 1 (12)

where i ∈ { 1,2 }. The spatial Softmax function enforces that the sum of attention weights across all pixels equals 1. This competitive probability distribution forces the network to allocate the extremely scarce high-weight values only to the most salient regions (i.e., the small targets), thereby ruthlessly suppressing large areas of invalid background (e.g., roads, bare land, and clouds).

3.5.3 Residual compensation and final fusion

To prevent the gradient vanishing of sparse small-target pixels during the attention masking process, we adopt a residual compensation structure. The generated spatial attention maps M i are element-wise multiplied with the original aligned features, and the initial features are added back, as formulated in Equation 13:

f i ′ = f i ⊗ M i + f i (13)

This residual design perfectly protects the original high-frequency details from being over-attenuated. Finally, the enhanced features f 1 ′ and f 2 ′ are concatenated along the channel dimension and passed through a 1 × 1 convolution to smooth the aliasing effect and generate the final output, as formulated in Equation 14:

F out = C o n v 1 × 1 C o n c a t f 1 ′ , f 2 ′ (14)

By replacing standard concatenation with CSAF, SOE-YOLO significantly enhances its robustness against complex remote sensing backgrounds while ensuring high localization precision for extremely small objects.

4 Experimental results and analysis

4.1 Datasets and experimental parameters

In this study, we use the VEDAI (Razakarivony and Jurie, 2016), RSOD (Long et al., 2017) and NWPU VHR-10 datasets (Cheng et al., 2014).

VEDAI consists of aerial imagery from agricultural areas in Utah, USA, focusing on vehicle detection in complex backgrounds. It provides both visible light (RGB) and near-infrared (NIR) multispectral images with a spatial resolution of approximately 12.5 cm/pixel. The dataset includes 1,210 images (sized 1024 × 1024 or 512 × 512) covering various scenes such as farmland, roads, and residential areas. Vehicle targets are extremely small, typically occupying only a few dozen pixels, and often exhibit similar colors and textures to the background.

RSOD is derived from high-resolution Google Earth satellite imagery and emphasizes multi-class object detection, particularly small and sparsely distributed targets. The RGB images have resolutions ranging from 0.3 to 2 m/pixel, with a total of 976 images of varying dimensions (from 300 × 300 to 1500 × 1500 pixels). The dataset annotates six object categories—aircraft, playgrounds, overpasses, oil tanks, vehicles, and ships—with 6,950 instances labeled as horizontal bounding boxes (HBB). Key challenges include significant scale variations, sparse samples in certain categories, and complex background interference.

NWPU VHR-10 is sourced from Google Earth and Vaihingen aerial imagery, containing RGB images at resolutions between 0.5 and 2 m/pixel. It comprises 800 images (650 with objects, 150 background), with sizes ranging from 500 × 500 to 1500 × 1500 pixels. The dataset includes annotations for 10 object categories—aircraft, ships, oil tanks, baseball fields, tennis courts, basketball courts, athletic fields, harbors, bridges, and vehicles—totaling 3,250 HBB-annotated instances. It features densely distributed targets and high inter-class similarity among certain categories.

Following the widely adopted COCO criterion (Lin et al., 2014), we define objects with bounding box areas smaller than 32 × 32 pixels as small objects, objects with areas between 32 × 32 and 96 × 96 pixels as medium objects, and objects with areas larger than 96 × 96 pixels as large objects. Given the diverse spatial resolutions of remote sensing datasets, we further adopt a relative scale criterion: targets whose bounding box area occupies less than 1% of the total image area are also classified as small objects, as their effective feature representations are equally limited by the extremely low pixel coverage.

In this paper, the original dataset preset ratio is used to divide the training, validation, and test sets. For hardware setup, the experiments in this paper are done based on 20.04 server with Intel Core i9-11900KF, Nvidia GeForce RTX 2080Ti 11G VRAM and 32 GB RAM. We used PyTorch 1.10.0 and Torchvision 0.11.1 as deep learning frameworks, improved based on Ultralytics YOLO version 8.3.0, and trained with official pre-training weights. Table 2 details the main parameter settings used during training.

TABLE 2. The main parameter settings.

All three datasets (VEDAI, RSOD, NWPU VHR-10) provide annotations in the Horizontal Bounding Box (HBB) format, and SOE-YOLO accordingly outputs HBB predictions. SOE-YOLO does not currently support Oriented Bounding Boxes (OBB), which would require additional angle regression heads and rotated anchor designs. While the DASI module’s dimension-aware adaptive fusion can partially accommodate multi-oriented targets by selectively weighting spatial and channel dimensions, the final detection head produces axis-aligned boxes. Extending SOE-YOLO to OBB detection—for instance, by incorporating a rotated detection head as in YOLOv8-OBB—is an important direction for future work, particularly for datasets containing densely packed, arbitrarily oriented targets (e.g., ships in harbors, vehicles in parking lots).

All experiments employ the standard Ultralytics YOLO augmentation pipeline, which includes Mosaic (4-image stitching with random cropping and scaling), Mixup (alpha-blending of two images with ratio λ ∼ B e t a ( 32.0 , 32.0 )), random horizontal/vertical flipping, HSV color jittering, and random scaling (0.5–1.5×). Notably, we did not modify any of these augmentations to specifically accommodate the HFCE module’s high-frequency priors. This is a deliberate design choice: the fixed-weight Center-Surround contrast filters in HFCE operate as zero-parameter high-pass operators that are invariant to photometric augmentations (color jittering, Mixup blending) because they respond to spatial frequency structure rather than absolute pixel intensity. Mosaic augmentation, which concatenates four random crops, naturally preserves the local high-frequency edge structure of individual targets within each crop quadrant. We empirically verified that disabling Mosaic during training degrades both the baseline YOLO11n (mAP50 drops by 1.8%) and SOE-YOLO (mAP50 drops by 2.1%), confirming that Mosaic is beneficial for small object augmentation irrespective of the HFCE prior. No augmentation parameters were tuned specifically for SOE-YOLO, ensuring a fair comparison with the baseline.

4.2 Ablation experiment and comparative analysis

To validate the performance improvements of the modules designed in this study, we conducted ablation experiments on the VEDAI dataset. We used YOLO11n as the baseline without pre-trained model initialization, evaluating the impact of different modules on detection performance under identical experimental conditions. The ablation experiment results are presented in Table 3.

TABLE 3. Ablation experiments.

As shown in Table 3, we evaluated model performance using metrics including mAP50, mAP50:95, GFLOPs, and parameter count. Each improvement experiment demonstrated performance enhancement, which we analyze below.

The first experimental group represents the baseline model using YOLO11n, achieving an mAP50 of 60.4%, mAP50:95 of 35.4%, with 3.007M parameters and a computational complexity of 8.1 GFLOPs. After introducing the Asymmetric Padding Downsampling Module (APDS), mAP50 increased by 2.3% from 60.4% to 62.7%, and mAP50:95 improved by 2.2% from 35.4% to 37.6%. This demonstrates that APDS effectively addresses feature misalignment issues while preserving more edge details in different directions through asymmetric padding, reducing information loss during downsampling, and enhancing small object detection performance. Notably, parameter count and computational complexity remained virtually unchanged from the baseline, achieving accuracy improvements without additional cost.

The third experimental group introduced C3k2-HFCE to the baseline model, resulting in an mAP50 increase of 2% from 60.4% to 62.4%, and mAP50:95 improvement of 1.8% from 35.4% to 37.2%. Analysis of per-class accuracy revealed improvements in small object categories such as cars and vans, confirming that the C3k2-HFCE enhanced small object detection and localization precision.

The fourth experimental group incorporated the DASI module, yielding an mAP50 improvement of 0.8% and an mAP50:95 increase of 1.1% (from 35.4% to 36.5%). This indicates that the module preserved small object details while leveraging semantic information from higher-level features to enhance small object representations, thereby improving the model’s localization capability.

The fifth experimental group implemented CSAF, which, compared to the baseline, increased mAP50 by 2.0% and mAP50:95 by 2.3%, with minimal increases in parameters and virtually unchanged GFLOPs. Remarkably, these significant accuracy gains are achieved with only a marginal increase in model parameters and virtually unchanged computational overhead (GFLOPs). The notable 2.3% increase in mAP50:95 specifically highlights the CSAF’s capability in high-precision localization. By employing cross-level semantic recalibration, CSAF effectively aligns high-level semantic cues with low-level spatial details, preventing the weak features of tiny targets from being drowned out. Furthermore, the competitive spatial Softmax attention ruthlessly suppresses complex background clutter (e.g., roads and clouds), while the residual compensation flawlessly preserves the high-frequency boundaries of small objects, leading to a highly robust and efficient detection neck.

The sixth, seventh, and eighth experimental groups combined the aforementioned modules, with the eighth group representing our final proposed model. Compared to the baseline, this model achieved a 7.4% improvement in mAP50 and a 4.4% increase in mAP50:95, demonstrating significantly enhanced detection and localization performance.

To quantify the synergistic relationships among modules, we compute the coupling gain for each module pair and the full combination. Let Δ i denote the independent gain of module i (measured as the mAP50 improvement over baseline when applied alone, from Experiments 2–5), and Δ i , j denote the gain when modules i and j are jointly applied (from Experiments 6–7). The interaction effect is defined as δ i j = Δ i , j − ( Δ i + Δ j ). A positive δ i j indicates super-additive synergy (the modules reinforce each other); a negative value indicates sub-additive redundancy (the modules address overlapping issues).

As shown in Table 4, the APDS + HFCE pair exhibits near-zero interaction ( δ = 0.0 ), indicating that spatial preservation and high-frequency enhancement address orthogonal problems—APDS corrects geometric misalignment while HFCE amplifies contrast, with minimal functional overlap. Adding DASI to APDS + HFCE yields a notably positive interaction ( δ = + 1.3 ), suggesting that DASI’s adaptive multi-scale fusion benefits synergistically from the cleaner, better-aligned features produced by APDS and the enhanced high-frequency signals from HFCE. The full four-module combination achieves a modest but positive interaction ( δ = + 0.3 ), confirming that while each module’s primary contribution targets a distinct bottleneck, the super-additive effect is most pronounced in the three-module setup (APDS + HFCE + DASI), whereas CSAF’s background suppression provides additional orthogonal benefit that largely stacks linearly on top of the preceding three modules.

TABLE 4. Cross-module coupling analysis on the VEDAI dataset. Δ i = independent gain; Δ i , j = joint gain; δ i j = interaction effect.

Bold values indicate the best performance or optimal results for each evaluation metric across the compared models or configurations (e.g., the highest mean Average Precision [mAP] and Frames Per Second [FPS], or the fewest parameters [Params] and computation costs [GFLOPs]).

4.3 Comparative experiments

To comprehensively evaluate the effectiveness of the proposed model in this study, a comparative analysis was conducted across three publicly available remote sensing object detection datasets (NWPU VHR-10, RSOD, and VEDAI) against several state-of-the-art models, including SuperYOLO (Zhang et al., 2023), YOLO-Drone (Zhai et al., 2023), SP-YOLOv8 (Ma and Pang, 2023), YOLO-TUF (Chen et al., 2024), OD-YOLO (Chen et al., 2024), YOLORemote (Bai et al., 2025), and LAR-YOLOv8 (Yi et al., 2024). The results detailed in Table 5, were primarily analyzed along two dimensions: model complexity and detection accuracy.

TABLE 5. Performance comparison of different object detection models on remote sensing datasets. FPS is measured on NVIDIA RTX 2080Ti with batch size one at 512 × 512 input resolution.

Bold values indicate the best performance or optimal results for each evaluation metric across the compared models or configurations (e.g., the highest mean Average Precision [mAP] and Frames Per Second [FPS], or the fewest parameters [Params] and computation costs [GFLOPs]).

In terms of model complexity, our model demonstrates a significant advantage in lightweight design. The total number of parameters (Params) is only 3.357 million (corresponding to a model file size of approximately 6.8 MB under FP32 storage; note that parameter count and model file size are fundamentally different metrics—the former denotes the number of learnable weights, while the latter reflects the storage volume after serialization), the smallest among all compared models. Compared to the larger models such as LAR-YOLOv8 with 26.37 million parameters and SP-YOLOv8 with 10.24 million, parameter reduction amounts to approximately 87.6% and 68.2%, respectively. The computational load (FLOPs) of our model is 10.8 G, ranking among the most efficient, comparable to YOLORemote (8.6 G) and LAR-YOLOv8 (9.4 G), and significantly lower than the highest-consuming YOLO-TUF at 48.2 GFLOPs.

Regarding detection performance, our model maintains high efficiency while achieving highly competitive comprehensive metrics. Specifically, on the NWPU VHR-10 dataset, it attained the highest mean Average Precision (mAP50) of 89.5%, indicating robust recognition capabilities across diverse remote sensing targets. On the RSOD dataset, although YOLO-TUF and YOLORemote both achieved the highest mAP50 of 92.8%, our proposed model outperformed all others in mAP50:95, with a value of 63.8%, reflecting more precise localization predictions. Notably, on the VEDAI dataset, the proposed model demonstrated superior performance, with mAP50 at 67.8% and mAP50:95 at 39.8%, outperforming others such as OD-YOLO by 2.9% and 3.8% respectively. These results underscore the model’s enhanced generalization ability and robustness in challenging scenarios involving small targets and complex backgrounds.

From the perspective of inference speed, our model achieves the highest FPS of 76.3 among all compared models, surpassing the next fastest model YOLORemote (72.4 FPS) by 5.4% and outperforming similarly compact models such as LAR-YOLOv8 (68.9 FPS) by 10.7%. More significantly, compared with models achieving competitive detection accuracy, SOE-YOLO delivers 2.4× the inference speed of SP-YOLOv8 (31.8 FPS) and 4.1× that of YOLO-TUF (18.6 FPS), while simultaneously achieving superior mAP50. This exceptionally high frame rate (76.3 FPS, corresponding to 13.1 m per frame) makes SOE-YOLO suitable for real-time deployment on edge devices, comfortably exceeding the 30 FPS threshold widely accepted as real-time performance in practical applications. The combination of minimal parameters (3.357M), moderate FLOPs (10.8G), and top-ranking FPS confirms that SOE-YOLO achieves a genuine and well-balanced lightweight design, where architectural efficiency translates directly into real-world inference speed, not merely into theoretical parameter reduction.

Overall, the experimental findings clearly indicate that model performance does not correlate straightforwardly with parameter count or computational burden. For instance, the high computational YOLO-TUF did not achieve the best accuracy across all datasets, whereas our model, with minimal parameters and lowest computational load, still attained high detection accuracy. This robust performance validates the effectiveness of our architectural design and demonstrates a successful balance between model efficiency and detection precision.

To more directly illustrate the trade-off between efficiency and accuracy, Figure 10 presents the Pareto front analysis comparing SOE-YOLO with other lightweight models on the VEDAI dataset, where the x-axis represents computational cost (GFLOPs) and the y-axis represents detection accuracy (mAP50). An ideal model should reside in the upper-left corner (high accuracy, low computational cost). As shown, SOE-YOLO occupies the Pareto-optimal position, achieving the highest mAP50 (67.8%) among all compared models while maintaining a competitive computational cost of 10.8 GFLOPs. Notably, models with similar or lower GFLOPs (e.g., YOLORemote at 8.6 G, LAR-YOLOv8 at 9.4 G) achieve substantially lower mAP50 values (61.7% and 62.6%, respectively), while models with comparable accuracy (e.g., SP-YOLOv8 at 61.0% mAP50) require nearly 3× the computational cost (30.2 G). This confirms that SOE-YOLO achieves a superior efficiency–accuracy trade-off, making it particularly suitable for deployment on resource-constrained platforms such as UAVs.

4.4 Detection performance on different object categories

To validate the generalization performance of our proposed algorithm on other remote sensing datasets, experiments were conducted on the Remote Sensing Object Detection Dataset and the Northwestern Polytechnical University Very High Resolution-10 dataset. On the Remote Sensing Object Detection Dataset, our algorithm was compared with YOLO11n, and the results are presented in Table 6.

TABLE 6. Detection performance of YOLO11n and SOE-YOLO on the remote sensing object detection dataset.

As shown in Table 6, there are significant improvements in AP50 and AP95 across all categories. The mAP50 and mAP95 increased by 1.8% and 4.5% respectively. This indicates a notable enhancement in the model’s target localization performance. For the aircraft category, which consists of small targets, AP50 improved by 0.6% and AP95 improved by 2.5%. This demonstrates that the improved model significantly enhances the detection performance of small targets.

On the Northwestern Polytechnical University Very High Resolution-10 dataset, as showed in Table 7, our proposed algorithm demonstrates significant improvements over YOLO11n across all categories, with mAP50 increasing by 3.7% and mAP95 by 4.5%. These results highlight the enhanced detection capabilities of our model.

TABLE 7. Detection performance of YOLO11n and SOE-YOLO on the northwestern polytechnical university very high Resolution-10 dataset.

4.5 Scale-specific detection performance

To further verify the effectiveness of the proposed method for small object detection, we conducted scale-specific performance evaluation on all three datasets. Following the COCO evaluation protocol (Lin et al., 2014), we partitioned all test instances into three scale categories—small (area < 3 2 2 pixels), medium (3 2 2–9 6 2 pixels), and large (> 9 6 2 pixels)—and computed AP50 and AP50:95 for each category separately. This scale-specific analysis is essential for rigorously evaluating small object detection methods, as aggregate mAP metrics can obscure significant performance variations across different target scales (Zhang et al., 2026a). The results are presented in Tables 8–10.

TABLE 8. Detection performance of YOLO11n and SOE-YOLO on the VEDAI dataset across different object scales.

Bold values indicate the best performance or optimal results for each evaluation metric across the compared models or configurations (e.g., the highest mean Average Precision [mAP] and Frames Per Second [FPS], or the fewest parameters [Params] and computation costs [GFLOPs]).

TABLE 9. Detection performance of YOLO11n and SOE-YOLO on the RSOD dataset across different object scales.

Bold values indicate the best performance or optimal results for each evaluation metric across the compared models or configurations (e.g., the highest mean Average Precision [mAP] and Frames Per Second [FPS], or the fewest parameters [Params] and computation costs [GFLOPs]).

TABLE 10. Detection performance of YOLO11n and SOE-YOLO on the NWPU VHR-10 dataset across different object scales.

Bold values indicate the best performance or optimal results for each evaluation metric across the compared models or configurations (e.g., the highest mean Average Precision [mAP] and Frames Per Second [FPS], or the fewest parameters [Params] and computation costs [GFLOPs]).

As shown in Table 8, SOE-YOLO achieves the most significant improvement on small-scale objects in the VEDAI dataset, with an AP50 improvement of 8.8% and an AP50:95 improvement of 5.7% compared to YOLO11n. The APDS module mitigates spatial feature misalignment during downsampling, directly preserving the fragile positional cues of small targets; the C3k2-HFCE module explicitly enhances high-frequency edge details that would otherwise be smoothed out by successive convolutions; and the CSAF module prevents the weak feature responses of tiny targets from being overwhelmed by background noise through cross-level semantic alignment and competitive spatial attention. These three modules collectively contribute the majority of the small object gains. In contrast, the improvement on large objects is marginal (+1.1% AP50), as the baseline model already achieves near-saturated performance for targets with sufficient pixel coverage.

Tables 9, 10 confirm that this pattern generalizes to the RSOD and NWPU VHR-10 datasets. On RSOD, the small-object AP50 improvement (+3.5%) is moderate compared to VEDAI, partly because the small aircraft targets in RSOD already achieve a relatively high baseline AP50 of 80.0%. However, the AP50:95 improvement for small objects reaches 5.4%, indicating that SOE-YOLO substantially enhances localization precision for small targets even when recognition accuracy is already high. On NWPU VHR-10, small objects show a pronounced AP50 gain of 6.5%, driven primarily by the vehicle category (AP50 improvement of 7.1%, Table 7) which is the most challenging small-object class in this dataset. Across all three datasets, the AP50:95 improvements consistently exceed the AP50 improvements for small objects, underscoring that the proposed modules are particularly effective at refining bounding box localization for small targets rather than merely improving classification.

Interestingly, the improvements on medium and large objects are not negligible. On RSOD, large objects (overpasses and playgrounds) show a 1.5% AP50 gain, attributable to the DASI module’s dimension-aware adaptive fusion that handles multi-oriented and elongated targets. On NWPU VHR-10, the harbor category (a large but structurally complex target) benefits substantially from CSAF’s background suppression, yielding a 12.1% AP50:95 improvement (Table 7). These results demonstrate that while SOE-YOLO is primarily designed for small object enhancement, its architectural innovations also confer measurable benefits across all object scales.

4.6 Visualization experiments

4.6.1 Grad-CAM feature visualization

To provide intuitive evidence that the CSAF module effectively preserves small-target feature responses that are otherwise drowned out in standard YOLO11, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the feature attention maps of both models. Figure 11 presents Grad-CAM comparisons, where multiple small aircraft targets are distributed across complex airport backgrounds including runways, tarmac, and surrounding vegetation.

As shown in Figure 11a, for standard YOLO11n, the Grad-CAM heatmaps reveal a severe attention dilution problem. In the first scene (top), the network’s attention is broadly activated across the entire tarmac and runway surface, forming a large diffuse yellow region that indiscriminately covers both the aircraft targets and the irrelevant concrete background. In the second scene (bottom), a similar pattern emerges: the Grad-CAM response spreads across large portions of the airport ground, with the individual aircraft receiving only marginally stronger activation than the surrounding pavement. Taken together, these visualizations quantitatively confirm the “drowning out” phenomenon hypothesized in Section 1: the weak feature responses of small targets, after undergoing multiple downsampling stages, are overwhelmed by the dominant background features during naive concatenation in the FPN-PAN neck. The network essentially “sees” the background context rather than the targets themselves.

In stark contrast, as shown in Figure 11b, SOE-YOLO’s Grad-CAM heatmaps exhibit dramatically different activation patterns. In both scenes, the attention maps display highly concentrated and spatially precise hotspots exclusively on each individual aircraft target. The intervening background regions (runway surfaces, tarmac, surrounding vegetation) receive near-zero activation, appearing as dark blue regions in the heatmap. The first scene shows discrete, tight red hotspots aligned with each parked aircraft, while the second scene reveals a similar pattern of sharply localized attention despite the more cluttered airport layout. This qualitative difference directly validates the effectiveness of CSAF’s two core mechanisms: the cross-level semantic recalibration bridges the semantic gap between shallow and deep features by establishing a cross-channel correlation matrix, enabling the model to recognize small targets as semantically meaningful entities rather than background noise; and the competitive spatial Softmax attention enforces a strict probability distribution across the spatial dimension, ruthlessly allocating the scarcest attention weights exclusively to the most salient small targets while aggressively suppressing expansive background clutter. The residual compensation path further ensures that the high-frequency boundaries of small targets are preserved during this attention redistribution process. These Grad-CAM results provide compelling visual evidence that CSAF fundamentally transforms the network’s attention mechanism from diffuse background-dominated activation to precise target-focused activation.

4.6.2 Detection result visualization

Figure 12 presents selected detection results of the algorithm proposed in this paper on the VEDAI test set; the first line in Figure 12 is the original image, the second line shows the detection outcomes of the original YOLO11n′s algorithm, and the third line illustrates the SOE-YOLO algorithm proposed in this paper. Comparing the detection results of the original YOLO11 algorithm with the algorithm proposed in this paper, in Group 1 the target located in the upper-right corner of the image is a car, the YOLO11n algorithm incorrectly classifies target as a pickup, whereas SOE-YOLO correctly identifies it. Furthermore, compared to YOLO11n, SOE-YOLO not only correctly identifies all the targets in this image, it also enhances the confidence level for each target. In Group 2, SOE-YOLO correctly recognizes more targets. In Group 3, YOLO11 incorrectly identifies the target as pickup and SOE-YOLO correctly identifies it as car. In Group 4, YOLO11 incorrectly identifies the target as pickup and SOE-YOLO correctly identifies it as car. In Group 5, SOE-YOLO improves the confidence level of the target.

Figure 13 shows some of the detection results of the algorithm proposed in this paper on the RSOD dataset. Comparing the detection results of the original YOLO11 algorithm with the algorithm proposed in this paper, in Group 1, both YOLO11n and SOE-YOLO correctly recognize all airplanes, and SOE-YOLO correctly recognizes all the targets in this image compared to YOLO11n, which improves the confidence level of each target. In Group 2 and Group 3, YOLO11n incorrectly recognizes other buildings or shadows as oiltank, which causes a misclassification; SOE-YOLO correctly recognizes other buildings or shadows as background. In Group 4, YOLO11 incorrectly identifies the building as an airplane, SOE-YOLO correctly identifies it as a background and more accurately locates the overpass. In Group 5, the location of the overpass is more accurately localized. In Group 6, SOE-YOLO detects the corner of the playground at the edge of the image.

Figure 14 shows some of the detection results of the algorithm proposed in this paper on the NWPU-VHR 10 dataset. Comparing the detection results of the original YOLO11 algorithm and the proposed algorithm in this paper, in Group 1, SOE-YOLO correctly recognizes the airplanes that are misrecognized by YOLO11, which reduces the misclassification. In Group 2, SOE-YOLO correctly identifies all targets in this picture, improving the confidence level of each target. In Group 3, SOE-YOLO correctly identifies the bridge and recognizes the ship that is not recognized by YOLO11n. In Group 4, compared with YOLO11n, SOE-YOLO correctly identifies all tennisball courts in this image and reduces the misidentification of car as an airplane. In Group 5, SOE-YOLO reduces the misclassification that recognizes the shadow as a car. In Group 6, SOE-YOLO reduces the misclassification of identifying clouds as harbor.

5 Conclusion

Due to the extreme scale variations, arbitrary orientations, and complex background interference inherent in remote sensing imagery, existing object detection networks often struggle with feature misalignment, information loss, and semantic gaps. These limitations inevitably lead to missed detections and localization biases, particularly for small targets. To address these critical challenges, this paper proposes SOE-YOLO (Small Object Enhanced-YOLO11), an efficient and accurate object detection framework. The proposed model introduces four key architectural innovations: (1) An Asymmetric Padding Down-Sampling (APDS) module mitigates spatial feature misalignment and preserves multi-directional edge details through position-aware directional padding and multi-branch fusion. (2) A High-Frequency Contrast Enhancement (HFCE) module, seamlessly integrated into the bottleneck architecture (C3k2-HFCE), acts as an inductive high-pass filter using parameter-free physical priors to explicitly highlight the high-frequency boundaries and local contrast of tiny targets. (3) A Dimension-Aware Selective Integration (DASI) module dynamically fuses multi-scale features via adaptive weight allocation, significantly enhancing multi-orientation adaptation and complex background suppression. (4) A Contextual Small-object Attention Fusion (CSAF) module replaces standard concatenation in the network neck, bridging semantic gaps through cross-level channel correlation and strongly suppressing cluttered backgrounds using a competitive spatial Softmax attention mechanism.

Extensive experiments on three public remote sensing datasets demonstrate the superiority of the proposed SOE-YOLO framework. On the VEDAI, RSOD, and NWPU VHR-10 datasets, the model achieves mAP50 scores of 67.8%, 90.8%, and 89.4%, respectively, representing absolute improvements of 7.4%, 1.9%, and 3.7% over the baseline YOLO11n model. Scale-specific evaluation further reveals that the improvements are most pronounced for small objects (bounding box area < 3 2 2 pixels): AP50 gains of 8.8%, 3.5%, and 6.5% on VEDAI, RSOD, and NWPU VHR-10, respectively, confirming that the proposed modules are particularly effective for enhancing small object detection. Notably, the AP50:95 improvements for small objects (5.7%, 5.4%, and 5.6% on the three datasets) consistently exceed the corresponding AP50 gains, underscoring the model’s strong capability in refining bounding box localization for tiny targets. Crucially, these accuracy gains are attained without compromising computational efficiency. The model maintains a highly lightweight profile with only 3.36 million learnable parameters (yielding a model file size of 6.8 MB under FP32 storage) and a computational cost of 10.8 GFLOPs. Because network parameters and model sizes are often strictly constrained in real-world engineering applications, achieving this state-of-the-art trade-off between detection accuracy and deployment efficiency holds significant practical value.

Despite these advancements, there remains room for further improvement. Future research will focus on expanding the training datasets with diverse samples and advanced data augmentation strategies to further improve the detection robustness of extremely small targets. To facilitate real-time edge deployment on platforms such as unmanned aerial vehicles (UAVs), advanced model compression techniques—including network pruning and knowledge distillation—will be explored to accelerate inference speed while maintaining high precision. Additionally, while this study validates the effectiveness of SOE-YOLO on standard RGB optical satellite and aerial imagery, its performance on significantly different modalities—such as Synthetic Aperture Radar (SAR), nighttime infrared imagery, or severely degraded low-resolution sources—remains an unexplored area. We plan to systematically evaluate the model’s cross-domain generalization capabilities and explore modality-specific adaptation strategies to broaden its practical applicability in complex, multi-source remote sensing scenarios. Furthermore, SOE-YOLO currently operates exclusively with Horizontal Bounding Boxes (HBB) and does not support Oriented Bounding Boxes (OBB). For remote sensing scenes with densely packed, arbitrarily oriented targets (e.g., ships in harbors, aircraft in parking areas), OBB detection can provide tighter localization and reduce overlap between adjacent instances. Incorporating an OBB detection head into SOE-YOLO is an important future direction that would significantly broaden its applicability.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Tianwen Zhang, Southwest Jiaotong University, China

Reviewed by: Xiaowo Xu, University of Electronic Science and Technology of China, China

Xiao Ke, University of Liverpool, United Kingdom

Zhongzhen Sun, National University of Defense Technology College of Electronic Engineering, China