GSDD-net: an efficient feature extraction method for gear-surface defects in complex industrial environments

1 Introduction

Growing demands for precision, load capacity, reliability, and intelligent operation in modern industrial equipment have increased the structural complexity of mechanical transmission systems and the severity of their operating conditions (Lv et al., 2022). Gears are essential transmission components widely used in automotive transmissions, aerospace gearboxes, marine propulsion systems, rail transit, heavy machinery, and advanced intelligent manufacturing equipment (Feng et al., 2023). Gear condition directly affects transmission accuracy, power transmission efficiency, system reliability, and service life. Gear defects include fractures (break), missing teeth (lack), scratches (scratch), chipping (kp), tooth-surface black-skin defects (hp_cm), and tooth-root black-skin defects (hp_cd). During prolonged operation, cyclic loading, impacts and vibration, variable lubrication conditions, and environmental factors can contribute to gear damage (Liu D. et al., 2023). Without timely and accurate detection, minor defects may progress under demanding operating conditions, including high loads and rotational speeds. Progressive damage can reduce transmission accuracy, increase noise and vibration, and accelerate component wear. In severe cases, it may cause sudden gear failure, unplanned equipment downtime, or catastrophic system failure, posing serious safety risks and causing economic losses. Therefore, developing accurate and robust methods for detecting gear surface defects is essential for ensuring safe and reliable equipment operation, improving quality control in intelligent manufacturing, and supporting predictive maintenance of critical transmission components.

Traditional gear inspection relied on visual examination, subjective judgment, and basic dimensional measurements to identify surface defects. These methods can be inefficient and susceptible to operator subjectivity. Their performance depends on operator experience, fatigue, and workplace conditions, making it difficult to meet the stringent quality control requirements of large-scale manufacturing (Lu W. et al., 2024). Sensor-based signal processing techniques are widely used to improve the efficiency of gear condition monitoring and fault diagnosis. For example, Xiao et al. (2024) detected gear faults by analyzing acoustic signals amplified by phononic crystal resonators. Another study proposed a spectral kurtosis-based blind deconvolution method combined with order-domain filtering for gear fault diagnosis (Hashim and Shakya, 2023). However, fault feature extraction using these techniques remains sensitive to environmental noise, variations in operating conditions, and sensor placement. These sensitivities can limit their reliability and adaptability in complex industrial environments. Advances in machine vision have made vision-based defect detection a promising approach to addressing some limitations of conventional inspection methods. This approach offers non-contact imaging, direct visualization of surface defects, and automated feature extraction (Yao et al., 2022).

Image-based gear defect detection methods can be broadly classified into three categories: traditional image processing, conventional machine learning, and deep learning. Traditional image processing methods rely primarily on handcrafted visual features and predefined rules to detect defects. Common techniques include edge detection, template matching, local binary pattern (LBP) feature extraction, threshold segmentation based on grayscale histograms, and mathematical morphology (Kong et al., 2022; So et al., 2013; Zhang et al., 2021; Win et al., 2015; Truong and Kim, 2018). Edge detection can efficiently delineate defect boundaries using gradient-based filters, but it is sensitive to noise. Template matching is suitable for detecting defects with consistent shapes but can be sensitive to variations in scale and orientation, as well as background clutter. LBP captures local texture variations and can improve detection robustness against complex backgrounds, although it remains sensitive to noise. Threshold segmentation and morphological operations can be effective when defects exhibit strong contrast with the background. However, their detection performance may decline for defects with varying scales, low contrast, and complex shapes. Overall, reliance on handcrafted features and fixed rules makes traditional image processing methods vulnerable to background clutter, noise, and variations in defect shape in complex industrial environments. These limitations can hinder generalization and detection accuracy, making it difficult to meet the requirements of high-accuracy industrial inspection.

To address the reliance of conventional image processing techniques on handcrafted features and rigid regulations, machine learning approaches have increasingly been integrated into flaw identification and implemented throughout diverse industrial settings. These methods typically extract visual features, often using handcrafted descriptors, and use trained classifiers to identify defect categories. Some frameworks also incorporate defect localization. Various detection frameworks have been developed for specific industrial applications. For example, Zhang et al. (2018) developed an automated method for railway track defect detection that demonstrated effective detection and segmentation on noisy datasets and field images. Kong et al. (2019) proposed a framework combining the whale optimization algorithm with a support vector machine to assess tool wear during Ti-6Al-4V machining under varying operating conditions. German et al. (2012) proposed an image-based method for detecting surface defects in reinforced concrete columns and quantifying their characteristics. Xu et al. (2018) developed an RNAMlet method based on multiscale geometric analysis for detecting surface defects in steel sheets. Quiroz et al. (2018) developed a random forest-based method for detecting broken rotor bar faults in industrial machinery condition monitoring and maintenance. Krummenacher et al. (2017) proposed two machine learning systems for automatically detecting cracks, peeling, and out-of-roundness in railway vehicle wheels. Gu et al. (2021) proposed a machine vision-based method for automatically detecting and identifying bearing surface defects. Naddaf-Sh et al. (2023) proposed a machine learning-based system for detecting and grading defective stud welds. Yang et al. (2025) developed a system for classifying automotive body surface defects, reporting a mean precision of 93% and a recognition rate of 100% for pseudo-defect classification. Although these studies have advanced industrial defect detection, generalization and detection accuracy can remain challenging in complex industrial settings, potentially limiting their suitability for high-accuracy inspection. Gear surface defect detection presents particular challenges, including substantial variation in defect size, loss of local detail, and interference from complex metallic backgrounds. Balancing detection accuracy and computational efficiency is also challenging.

In complex industrial environments, diverse defect morphologies, substantial background noise, and stringent accuracy requirements pose challenges for conventional machine learning methods. These challenges are compounded by the limited representational capacity of handcrafted features and constraints on model generalization. These limitations can hinder their ability to meet practical inspection requirements. Consequently, deep learning-based object detection has become a major research focus in industrial defect inspection. Deep learning models, particularly convolutional neural networks, learn hierarchical feature representations, from local defect details to high-level semantic information, through multiple layers of nonlinear transformations. This capability reduces reliance on handcrafted features and can improve detection performance and robustness across diverse defect types and complex backgrounds. Deep learning-based object detectors are commonly categorized as two-stage or one-stage methods. Two-stage detectors, such as R-CNN (Girshick et al., 2014), Fast R-CNN (Girshick, 2015), Faster R-CNN (Ren et al., 2016), and Mask R-CNN (He et al., 2017), first obtain region proposals and then classify them and refine their bounding boxes. Although two-stage detectors can achieve high detection accuracy, their computational cost and inference latency may limit their suitability for real-time industrial inspection. One-stage detectors, such as SSD (Liu et al., 2016), RetinaNet (Lin et al., 2017), and the YOLO family, predict object classes and bounding boxes directly from feature maps without a separate region proposal stage. This design generally supports faster inference and lower computational cost. The YOLO family seeks to balance detection accuracy, inference speed, and ease of deployment through improvements in network architecture, feature fusion, and training strategies. YOLO-based detectors are therefore widely used in industrial defect detection. YOLO variants from YOLOv5 to YOLOv12 (Khanam and Hussain, 2024a; Li et al., 2022; Wang et al., 2023; Zhang et al., 2025; Wang CY. et al., 2024; Wang A. et al., 2024; Khanam and Hussain, 2024b; Tian et al., 2026) have introduced improvements aimed at multiscale object detection, robustness to background clutter, and real-time inference. YOLOv13 (Lei et al., 2025) incorporates a hypergraph-enhanced architecture to improve feature representation. YOLO26 (Sapkota et al., 2025) adopts an end-to-end detection architecture that eliminates non-maximum suppression (NMS) post-processing. Reported improvements include small-object detection, on-device deployment efficiency, and CPU inference performance. Deep learning-based object detection, particularly one-stage methods such as YOLO, therefore provides a promising foundation for accurate, real-time gear surface defect detection in complex industrial environments.

Recently, YOLO-based architectures have gained extensive popularity for object localization and industrial flaw recognition due to ongoing innovations in neural network designs, multi-scale feature integration, and optimization techniques. Li et al. (2023) proposed EFD-YOLOv4 for steel plate surface defect detection and reported improved detection performance. Zhang et al. (2023) proposed a YOLOv5-based method for gear surface defect detection that suppresses interference and reduces false positives (FPs) and false negatives (FNs). Liu et al. (2025) proposed REPBB-YOLOv7 for continuous casting slab inspection, reporting robustness, high localization accuracy, and real-time detection performance. Li et al. (2024) modified YOLOv8 for object detection in terahertz (THz) images, reporting mAP improvements of 5.0% and 3.9% over the baseline on THz Dataset I and Dataset II, respectively. Lu M. et al. (2024) developed WSS-YOLO, a YOLOv8-based model for industrial steel defect detection. Wang et al. (2025) developed SDS-YOLO based on YOLOv11 to locate concrete vibration sites in complex environments. Zhao C. et al. (2023) proposed RDD-YOLO for steel surface defect detection. Shuai et al. (2024) developed SF-YOLO for metal gear end-face defect detection, reporting improved detection accuracy and speed for small defects and defects of varying sizes. Advances in feature enhancement and multitask learning across computer vision also offer useful insights for industrial defect detection. For example, Yu et al. (2024a) proposed a deep soft-thresholding feature separation module with a trainable thresholding mechanism. They reported that this mechanism suppresses noise and enhances relevant features, improving model stability in complex environments. Yu et al. (2024b) further showed that multitask learning can improve feature discriminability and model adaptability by sharing feature representations and jointly optimizing related objectives. These approaches offer methodological insights for addressing variations in defect size, loss of local detail, and interference from complex metallic backgrounds in gear surface defect detection. They may also inform efforts to balance detection accuracy and computational efficiency.

Recent studies have explored methods for addressing challenges in industrial surface defect detection, including variations in defect size, loss of local texture information, and interference from complex metallic backgrounds. Zhang et al. (2024) developed a YOLOv5-based detection architecture that enhances feature extraction and fusion through a multiscale feature fusion module and an attention-based residual module. The model achieved an mAP of 78.5% on the NEU-DET dataset, although the additional modules increased model complexity and reduced inference speed. Niu et al. (2020) proposed a surface defect generative adversarial network (SDGAN) that combines a dual-diversity control discriminator with a cycle-consistency loss to generate defect samples. This approach was reported to improve robustness to illumination changes. However, the method may require offline optimization to meet real-time monitoring requirements. Liu M. et al. (2024) proposed YOLOv8-LMG, incorporating VanillaNet, the Lion optimizer, and a CFP-EVC module to improve detection accuracy and inference speed. Liu B. et al. (2024) developed CAC-YOLOv8, which uses a channel attenuation network and a combined-pooling spatial pyramid module to enhance multiscale feature fusion and receptive-field adaptability. Zhao Y. et al. (2023) developed GRP-YOLOv5, incorporating gamma correction, the ResC2Net residual architecture, and partial convolution (PConv) to improve image preprocessing and feature extraction. In a related study on vibration-based gear fault diagnosis, Li et al. (2019) used particle swarm optimization (PSO) to optimize a deep neural network (DNN), achieving a classification accuracy above 90% using raw vibration data. However, detecting early-stage pitting and complex defects at multiple locations remains challenging.

Despite advances in gear defect detection, several challenges remain in complex industrial environments. Gear defects vary considerably in size and shape, and their local features can be obscured by complex metallic backgrounds. These characteristics complicate reliable detection and accurate localization. The computational and memory requirements of existing models can also hinder deployment on resource-constrained industrial devices. Balancing detection accuracy, inference speed, and robustness to varying operating conditions remains challenging, particularly for small gear defects or those with irregular shapes. Further improvements in overall detection performance are therefore needed.

To address the key challenges in gear surface defect detection, including diverse defect morphologies, limited spatial representation of local defects, spatial detail loss, complex background interference, and the trade-off between detection accuracy and real-time performance, this paper proposes GSDD-Net, an efficient YOLO-based framework for feature representation and multi-scale defect detection. The proposed framework aims to achieve accurate and reliable gear surface defect detection. The main contributions of this work are summarized as follows:

• To mitigate the loss of edge contours and texture details from subtle defects during feature downsampling, we propose the ADown module to replace the conventional downsampling operation. The module further reduces the computational cost introduced by convolution-based downsampling. ADown combines average pooling and max pooling to capture complementary global contextual cues and locally salient features with minimal computational overhead. Within GSDD-Net, ADown preserves weakly salient defect features while maintaining a low parameter count and computational cost.
• To enhance fine-grained feature representation, cross-channel interaction, and multi-scale feature fusion, we design the CSP-EDLAN module for the feature extraction and fusion stages of GSDD-Net. The module integrates cross-stage partial connections, multi-gradient path aggregation, and DualConv-based feature extraction. In GSDD-Net, CSP-EDLAN reduce redundant computation while facilitating gradient propagation, cross-channel interaction, and multi-scale feature aggregation for diverse defect patterns.
• To alleviate the effects of uneven illumination, oil contamination, and metallic reflections, we propose the C2TSSA module for global context modeling. Based on token statistical self-attention, the module derives attention weights from token-level statistics and performs global feature modeling with linear computational complexity. Within GSDD-Net, C2TSSA enhances the representation of weakly salient regions while suppressing irrelevant background responses.
• To alleviate the mismatch between classification confidence and bounding-box localization quality in decoupled detection heads, we design LQEHead based on localization quality estimation. LQEHead extracts localization stability features from bounding-box distribution predictions and uses them to calibrate classification logits before the Sigmoid operation. This design enables the final confidence score to jointly reflect category evidence and localization reliability.

The remainder of this paper is organized as follows. Section 2 introduces the overall technical framework and the proposed algorithm. Section 3 analyzes and discusses the experimental results. Section 4 summarizes the experimental findings.

2 The proposed methods

To satisfy the demands for lightweight and high-precision detection of gear surface flaws in intricate industrial settings, this research presents GSDD-Net, an efficient feature extraction and multi-scale identification framework grounded in the YOLO structure, as illustrated in Figure 1. Utilizing gear defect photographs as inputs, the model addresses four practical requirements: fine-grained feature retention, multi-scale semantic integration, complex background suppression, and localization quality awareness, thereby supporting automated recognition and localization of representative flaws, such as break, lack, scratch, kp, hp_cm, and hp_cd. Specifically, during the feature downsampling phase, the ADown module is incorporated to perform a streamlined reconstruction of the standard downsampling procedure. By merging average pooling and max pooling procedures, ADown reduces model complexity while retaining local edges, textures, and faintly salient defect information. Within the backbone feature extraction phase and the neck cross-scale fusion phase, the CSP-EDLAN module is incorporated to refine the feature representation units. By incorporating cross-stage partial connections, multi-gradient path aggregation, and the DualConv dual-convolution feature extraction mechanism, CSP-EDLAN reduces redundant computation while strengthening gradient transmission, channel interaction, and multi-scale feature aggregation for gear defects with diverse scales and morphologies. Moreover, the C2TSSA module is incorporated to model global contextual dependencies with linear complexity based on a token statistical self-attention mechanism, thereby providing contextual information for faintly salient defect areas and complex backgrounds such as metallic reflections, oil contamination, and uneven lighting. Ultimately, the localization quality estimation detection head, LQEHead, is engineered to extract localization stability features from bounding-box distribution forecasts and coordinate classification confidence with localization precision. In conclusion, the targeted integration of lightweight downsampling, feature representation, global context modeling, and a quality-aware detection head yields the reported precision, robustness, and inference efficiency under the tested settings while maintaining low model complexity. The resulting configuration provides a task-specific framework for gear surface flaw detection in complex industrial environments.

2.1 ADown module

Gear surface defects, such as break, lack, scratch, kp, hp_cm, and hp_cd, often occupy only a small portion of the input image, making the preservation of local edge and texture information important during downsampling. Conventional strided convolution can efficiently reduce spatial resolution, but may also discard fine-grained information that is useful for recognizing small or weak defects. To address this issue, GSDD-Net incorporates the ADown module into the downsampling path as a lightweight alternative to conventional convolution-based downsampling. ADown combines average pooling, max pooling, and convolutional operations to reduce spatial resolution while retaining complementary global and local feature responses with limited computational overhead. In GSDD-Net, ADown is therefore used as a task-oriented downsampling strategy to better preserve defect-related details, and its contribution to the overall detector is evaluated through ablation experiments.

As shown in Figure 2, ADown combines average pooling, max pooling, and convolution in a two-branch structure, with the input feature map X ∈ R C × H × W first split along the channel dimension into two sub-feature maps, collectively denoted as X 1 , X 2 ∈ R C 2 × H × W, and the specific calculation process is presented in Equation 1.where X ∈ R C × H × W denotes the input feature map with C channels, height H, and width W; X 1 and X 2 denote the two feature maps produced by splitting X into equal channel groups, each with C / 2 channels. Split denotes channel-wise splitting. The symbol R used in the tensor notation denotes the real-number space.

X 1 , X 2 = S p l i t X (1)

The first branch applies convolution to extract features, and the specific calculation process is presented in Equation 2.where F 2 denotes the output of the second branch, X 2 is the second split feature map, MaxPool denotes the max-pooling operation in this branch, and Conv denotes the subsequent convolution. The two branch outputs have the same spatial dimensions for concatenation.

F 1 = C o n v X 1 (2)

The two branch outputs are concatenated along the channel dimension, and the specific calculation process is presented in Equation 3.

F 2 = C o n v MaxPool X 2 (3)

Here, max pooling is performed on 2 × 2 with a stride of 2 to reduce spatial resolution while preserving salient features. The feature maps F 1 , F 2 ∈ R C 2 × H 2 × W 2 output by the two branches have the same spatial dimensions. Finally, the outputs of the two branches are concatenated along the channel dimension to obtain the downsampled feature map Y ∈ R C × H 2 × W 2, and the specific calculation process is presented in Equation 4.where Y denotes the downsampled output and Concat denotes concatenation of F 1 and F 2 along the channel dimension. With the channel dimensions used here, Y has C channels and spatial dimensions H / 2 × W / 2.

Y = C o n c a t F 1 , F 2 (4)

Compared with standard large-stride convolution, the ADown module combines pooling operations with a compact convolutional structure to perform spatial downsampling with relatively low computational overhead. The parallel branches retain complementary global and local feature responses, which is intended to reduce the loss of edge and texture information associated with small gear defects during resolution reduction. In GSDD-Net, ADown is used primarily as a lightweight downsampling component, while its effect on detection accuracy and model complexity is assessed through the ablation experiments.

2.2 CSP-EDLAN module

To improve feature representation and multi-scale feature fusion for gear surface defect detection, GSDD-Net incorporates a Cross-Stage Partial Efficient Dual Layer Aggregation Network (CSP-EDLAN) module into the backbone and neck of YOLO11, replacing the original C3K2 blocks. Rather than introducing each underlying operation as an independent novel design, CSP-EDLAN integrates cross-stage partial connections, multi-gradient path aggregation, and DualConv units within a unified feature extraction and fusion structure. This configuration is intended to promote feature reuse, facilitate gradient propagation and cross-channel interaction, and support the representation of gear defects with different scales and morphological characteristics while controlling computational redundancy. In GSDD-Net, CSP-EDLAN therefore serves as a task-oriented feature aggregation module for adapting the baseline network to the multi-scale characteristics of gear surface defects. The structure of the CSP-EDLAN module is shown in Figure 3.

The structure of CSP-EDLA is designed as follows. Given an input feature map I = R H × W × C i n, the module first adjusts the number of channels to C through a 1 × 1 convolutional layer (CBS):

CSP-EDLAN first adjusts the feature channels using a 1 × 1 convolutional layer, the specific calculation process is presented in Equation 5. where I denotes the input feature map of spatial size H × W and input channel count C i n; I ′ denotes the channel-adjusted feature map with C channels. The operator f c o n v 1 × 1 denotes the 1 × 1 convolution used in the CBS unit. CBS denotes convolution, batch normalization, and SiLU activation.

I ′ = f c o n v 1 × 1 I ∈ R H × W × C (5)

The resulting feature map is split into a shortcut branch and a feature-extraction branch, and the specific calculation process is presented in Equation 6.where I s h o r t c u t denotes the shortcut feature retained for later concatenation, I e l d a n denotes the feature supplied to EDLAN, and Split divides I ′ into two equal groups along the channel dimension. Each branch contains C / 2 channels.

I s h o r t c u t , I e d l a n = S p l i t I ′ (6)

EDLAN applies DualConv progressively, beginning with the first layer and then passing each layer output to the next, the specific calculation process is presented in Equations 7, 8.where Z 1 denotes the output of the first DualConv layer and I e l d a n is the input to that layer. DualConv denotes the convolutional unit specified in Equation 11.where Z j and Z j − 1 denote the outputs of the current and preceding DualConv layers, respectively; j is the layer index, and n is the total number of DualConv layers in EDLAN. The superscripts indicate layer indices rather than exponentiation.

Z 1 = D u a l C o n v I e d l a n (7)

Z j = D u a l C o n v Z j − 1 , j = 2 , 3 , . . . , n (8)

The shortcut feature and the intermediate DualConv outputs are concatenated along the channel dimension, and the specific calculation process is presented in Equation 9.where Z c o n c a t denotes the concatenated feature map, Concat denotes channel-wise concatenation, and Z 1 through Z n are the intermediate DualConv outputs. Concatenating these n outputs with I s h o r t c u t gives n + 1 C / 2 channels at spatial size H × W.

Z c o n c a t = C o n c a t I s h o r t c u t , Z 1 , Z 2 , . . . , Z n ∈ R H × W × n + 1 · C / 2 (9)

A final 1 × 1 convolution fuses the concatenated features and adjusts the output channel count, and the specific calculation process is presented in Equation 10.where O ∈ R H × W × C o u t denotes the output of CSP-EDLAN, C o u t denotes its output channel count, and f c o n v 1 × 1 denotes the final 1 × 1 convolution applied to Z c o n c a t for feature fusion and channel adjustment.

O = f c o n v 1 × 1 Z c o n c a t ∈ R H × W × C o u t (10)

DualConv combines a grouped spatial convolution branch and a pointwise convolution branch through element-wise addition, the specific calculation process is presented in Equation 11.where X denotes the input to DualConv, with height H, width W, and M input channels, and Y denotes its output. D W C o n v 3 × 3 G denotes the 3 × 3 convolution branch with G channel groups in this equation, while C o n v 1 × 1 denotes the 1 × 1 pointwise convolution branch applied to the full input. G is the number of groups, and + denotes element-wise addition of branch outputs with matching dimensions.

Y = D W C o n v 3 × 3 G X + C o n v 1 × 1 X (11)

In GSDD-Net, CSP-EDLAN is incorporated into the neck of YOLO11 to replace the original C3K2 block. The module combines cross-stage partial connections, multi-branch feature aggregation, and DualConv units to organize feature propagation and information interaction within the neck. Specifically, the cross-stage partial connection provides an additional route for feature reuse and gradient transmission, while the multi-branch structure of EDLAN aggregates features from different paths. In addition, the pointwise convolution branch in DualConv facilitates cross-channel information exchange, as described in Equation 11. Through this configuration, CSP-EDLAN is used to support feature extraction and multi-scale feature fusion in the neck, with particular consideration for gear defects with relatively small sizes or weak local characteristics, such as scratch, kp, hp_cm, and hp_cd.

2.3 C2TSSA module

Gear-surface defects, including break, lack, scratch, kp, hp_cm, and hp_cd, usually occupy only a small number of pixels and exhibit weak contrast against complex metallic backgrounds. These defects are easily affected by non-uniform illumination, grease, and metallic reflections. In the baseline YOLO11 (Khanam and Hussain, 2024b), the C2PSA block is located at the deepest stage of the backbone, after the SPPF module. Although C2PSA enhances feature representations through partial-channel attention and spatial-pyramid aggregation, its modeling of long-range spatial dependencies between defect regions and their surrounding context remains limited. Consequently, the responses of small and weakly salient defects may be suppressed by background interference.

To address this limitation, we replace the original C2PSA block with the proposed C2TSSA module at the same backbone position, as shown in Figure 1. C2TSSA uses token-statistical attention to capture contextual consistency across the feature map. It strengthens responses that are consistent with the surrounding gear structure while suppressing isolated responses caused by reflections and background textures. In this way, C2TSSA improves the representation of small and low-contrast defects without introducing explicit pairwise token interaction; its attention computation exhibits linear growth with respect to the number of tokens. The detailed structure of C2TSSA is illustrated in Figure 4.

Let the input feature be X ∈ R B × N × C, where B denotes the batch size, N denotes the number of tokens, and C denotes the channel dimension. The core idea of C2TSSA is to replace the pairwise token dot-product operation used in conventional self-attention with the construction of attention weights based on token-feature statistics. First, the mean vector μ ∈ R C and covariance matrix Σ ∈ R C × C of all tokens are computed, the specific calculation process is presented in Equation 12.

μ = 1 N ∑ i = 1 N X i , Σ = 1 N ∑ i = 1 N X i − μ X i − μ T (12)

Subsequently, the tokens are mapped to a low-dimensional semantic subspace through linear projection W r ∈ R C × d, and the statistical attention matrix A is constructed, the specific calculation process is presented in Equation 13.where W r is the linear projection matrix with C input dimensions and d projected dimensions, d is the projected feature dimension, Z denotes the projected token features, and A denotes the attention matrix obtained by Softmax normalization. The operator tr denotes the matrix trace, so t r Σ is the sum of the diagonal entries of the covariance matrix. λ is the regularization coefficient defined in Equation 14, and the centered dot denotes multiplication.

Z = X W r , A = S o f t ⁡ max Z Z T − λ · t r Σ (13)

The regularization coefficient is calculated as follows:where τ denotes the update step-size hyperparameter, p denotes the subspace dimension used to scale the covariance trace, and N is the number of tokens. The quantities ∑ and λ retain the meanings given below Equations 12, 13, respectively.

λ = − τ N · 1 1 + 1 p t r Σ (14)

The attention weights are applied to the input features to obtain the enhanced representation, as defined in Equation 15. The structure of TSSAttention is shown in Figure 5.where TSSA denotes the token statistical self-attention operation, T S S A X is the resulting feature representation, A is the attention matrix in Equation 13, X is the input token feature tensor, and the centered dot denotes matrix multiplication.

T S S A X = A · X (15)

Computational complexity comparison. The reported comparison between C2TSSA and the C2PSA attention branch is summarized in Table 1.

TABLE 1. Computational complexity comparison between the C2TSSA and C2PSA modules.

Where N denotes the token count, C the feature channel dimension, d the projected statistical dimension, and C a the number of channels processed by the C2PSA attention branch. The notation O · denotes asymptotic computational or memory cost. The expressions in Table 1 are stated per sample.

C2TSSA uses token statistical self-attention for global contextual modeling. Within GSDD-Net, it replaces the C2PSA block in YOLO11 to model contextual relationships between gear defect regions and their surroundings, with the aim of improving the representation of small defects.

2.4 LQEHead

In the baseline YOLO11, the decoupled detection head predicts classification logits and bounding-box distributions through separate branches. Although this design facilitates task-specific feature learning, the classification branch does not explicitly incorporate the localization quality estimated by the regression branch. Consequently, the classification confidence may not accurately reflect the localization accuracy of the predicted box. This inconsistency is particularly problematic for small gear-surface defects, such as break, lack, scratch, kp, hp_cm, and hp_cd, because even a slight displacement of a predicted box can result in a substantial decrease in overlap with the ground-truth box.

To address this issue, we replace the original YOLO11 detection head at the P3, P4, and P5 prediction scales with the proposed LQEHead. For each scale, LQEHead extracts localization-stability information from the predicted bounding-box distribution and uses it to generate a localization-quality correction. The correction is added to the classification logit before the Sigmoid operation, allowing the final confidence to reflect both category evidence and localization reliability. The high-resolution P3 branch is particularly beneficial for small defects, while the P4 and P5 branches provide stronger semantic representations for defects at other scales. This design helps preserve weak but accurately localized defect predictions and suppress poorly localized responses caused by background edges or metallic reflections. The structure of LQEHead is illustrated in Figure 6.

Let p ∈ R B × N × 4 denote the bounding-box distribution prediction output by the detection layer, where B denotes the batch size and N denotes the number of predicted boxes. Let s ∈ R B × N × C denote the initial classification score, and C denote the number of classes. LQEHead first extracts the localization-stability statistical feature f from the bounding-box distribution prediction, the specific calculation process is presented in Equation 16.

f = ϕ p = T o p K p , M e a n p (16)

The statistical feature f is subsequently mapped to a scalar localization-quality correction through a multilayer perceptron, the specific calculation process is presented in Equation 17.

q = M L P f (17)

Let s denote the raw classification logit generated by the classification branch. The quality correction q is fused with s through element-wise addition, the specific calculation process is presented in Equation 18.

s ′ = s + q (18)

Here, s ′ represents a quality-calibrated classification logit rather than a normalized probability. The final confidence score is obtained by applying the Sigmoid function, the specific calculation process is presented in Equation 19.

J = S i g m o i d s ′ (19)

The additive fusion is performed in the logit domain, allowing the localization-quality branch to provide a signed residual correction to the classification response. A positive correction increases the confidence of predictions with reliable localization, whereas a negative correction reduces the confidence of predictions with unstable bounding-box estimates. This formulation also avoids the excessive attenuation that may arise from directly multiplying two probability values, which is particularly important for small and low-contrast gear defects.

In the implementation, q is broadcast along the class dimension and added to the corresponding classification logits before the Sigmoid operation. The resulting score J is used as the final classification confidence during inference.

Overall, LQEHead incorporates localization quality assessment to better align bounding-box accuracy with classification confidence. Replacing the baseline detection head with LQEHead improves the localization and detection of small defects with minimal additional computational overhead, thereby reducing discrepancies between classification and regression outputs.

3 Experimental results

3.1 Evaluation metrics

To comprehensively evaluate GSDD-Net, both detection accuracy and computational efficiency were considered. The evaluation metrics included precision (P), recall (R), F1-score, class-wise average precision (AP), mean average precision (mAP), giga floating-point operations (GFLOPs), parameter count, and inference speed (FPS). Precision represents the proportion of predicted positive instances that are correct, whereas recall represents the proportion of ground-truth positive instances that are correctly detected. The F1-score is the harmonic mean of P and R, while AP corresponds to the area under the precision-recall (P-R) curve for a specific class. In the P-R curve, recall is plotted on the x-axis and precision on the y-axis, whereas mAP is calculated by averaging the AP values across all classes. GFLOPs quantifies computational complexity, parameter count reflects model size, and FPS indicates the number of images processed per unit time. The corresponding calculations are provided in Equations 20–25, where TP denotes correctly detected positive instances, FP denotes negative instances incorrectly detected as positive, and FN denotes positive instances missed by the model. In the FPS calculation, F n represents the number of images processed, and T represents the total time required to process them.

P = T P T P + F P (20)

R = T P T P + F N (21)

F 1 − s c o r e = 2 · P · R P + R (22)

A P = ∫ 0 1 P R d R (23)

m A P = 1 m ∑ i = 1 m A P i (24)

F P S = F n T (25)

3.2 Datasets and model training

GSDD-Net was evaluated separately on two publicly available gear defect datasets with distinct imaging conditions and defect categories. Each dataset defines a three-class detection task, and all three classes were included in model training, validation, testing, and class-wise performance evaluation. The training sets were used to train the model, whereas the validation and test sets were used to validate and test model performance, respectively. The model input size was set to 640 × 640 pixels for both datasets. Tables 2, 3 summarize the class-specific image counts.

• Actual Industrial Gear Defect Dataset (GID): The GID dataset was obtained from the publicly available industrial gear defect dataset released on the AI Studio platform (https://aistudio.baidu.com/datasetdetail/164152). It contains 2000 images collected in industrial manufacturing environments. The three defect classes are chipping (kp), tooth-surface black-skin defect (hp_cm)and tooth-root black-skin defect (hp_cd), corresponding to YOLO class IDs 0, 1, and 2, respectively. The images were annotated using LabelImg, and the XML annotation files were converted to the TXT format required by YOLO. The dataset was divided at a ratio of 7:2:1 into 1,400 training images, 400 validation images, and 200 test images. During preprocessing, all images were resized to 640 × 640 pixels for model input. Table 2 lists the number of images containing each class. Class-wise evaluation was conducted for hp_cm, hp_cd, and kp on the 200-image test set using AP50 and AP50-95; the corresponding P-R curves and confusion matrices are presented in Figures 7, 8.
• Public gear defect dataset: The public gear defect dataset was obtained from the Roboflow platform (https://universe.roboflow.com/gear-u48i0/gear-defect). It contains 2978 images with three defect classes: break, lack, and scratch, corresponding to YOLO class IDs 0, 1, and 2, respectively. Most images were acquired against a white background under a consistent viewing angle and uniform illumination. The dataset was partitioned at a ratio of 7:2:1 into 2084 training images, 595 validation images, and 299 test images. Table 3 reports the class-specific image. Because one validation image contains both lack and scratch, class-specific image counts are counted separately for the two classes. All three classes were used for model training, validation, testing, and class-wise evaluation. Figure 9 presents the P-R curves and AP50 values of YOLO11n and GSDD-Net for break, lack, and scratch on the 299-image test set.

TABLE 2. Class distribution of GID dataset.

TABLE 3. Class distribution of the public gear defect dataset.

The specific hardware and software setups utilized for the study are detailed in Table 4. The computational framework comprised an Intel Core i7-7265K central processing unit (CPU), an NVIDIA GeForce RTX 3090 24 GB graphics processing unit (GPU), 32 GB of DDR5 7200 MHz RAM, and a 1 TB M.2 NVMe solid-state drive, which collectively offered robust backing for extensive parallel computation and network training. The programming environment was built upon PyTorch 2.7.1, enhanced with CUDA 12.8, and executed under the Windows 11 platform to guarantee hardware-software integration and operational consistency. The training parameters are summarized in Table 5. The architecture was optimized using the SGD optimizer for weight adjustments, featuring an initial learning rate of 0.01. The momentum term and weight decay factor were configured at 0.937 and 0.0005, respectively. The total number of training epochs was fixed at 200, input frames were consistently scaled to 640 × 640 dimensions, and the batch size was adjusted to 32, which maintained a equilibrium between training velocity, processing demand, and GPU memory utilization. To ensure a fair comparison, all comparison models were trained under the same experimental settings described. These settings were derived from prior empirical research and further refined via grid search, thereby promoting consistent model convergence throughout extended training sessions. The learning-rate decay and warm-up protocols aided in stabilizing weight updates during the initial phases of training and enhanced the equilibrium and robustness of the optimization procedure throughout the following training epochs.

TABLE 4. Hardware and software environment.

TABLE 5. Training hyperparameter settings.

3.3 Ablation experiments

Table 6 presents the ablation results obtained on the GID dataset by incorporating different components into the YOLO11n baseline. The baseline model achieves 69.9% mAP50% and 35.9% mAP50-95, with a model size of 5.2 MB, 2.6 M parameters, and 6.3 GFLOPs. When introduced individually, CSP-EDLAN yields the largest accuracy improvement among the tested modules, increasing mAP50 and mAP50-95 by 1.5 and 1.0 percentage points, respectively. ADown mainly reduces model complexity, decreasing the model size to 4.3 MB and the parameter count to 2.1 M, while increasing mAP50 by 0.4 percentage points. C2TSSA and LQEHead also produce moderate improvements in the reported detection metrics, with LQEHead increasing mAP50-95 by 0.7 percentage points relative to the baseline. When multiple components are combined, the model shows further changes in detection performance. The combination of CSP-EDLAN and LQEHead achieves 72.1% mAP50% and 37.5% mAP50-95, corresponding to increases of 2.2 and 1.6 percentage points over the baseline, respectively. With ADown, CSP-EDLAN, and C2TSSA incorporated together, the model reaches 72.3% mAP50% and 37.9% mAP50-95 while reducing the parameter count to 1.9 M. After integrating all four components, GSDD-Net achieves 73.0% mAP50% and 38.0% mAP50-95. Compared with YOLO11n, this configuration improves mAP50 and mAP50-95 by 3.1 and 2.1 percentage points, respectively, while reducing the model size, parameter count, and GFLOPs by 23.1%, 26.9%, and 6.3%. Overall, the ablation results indicate that the integrated configuration provides a favorable balance between detection accuracy and model complexity under the experimental conditions of this study.

TABLE 6. Ablation experiments of each module in the proposed method.

To compare the performance of different feature enhancement modules, the original C2PSA module in YOLO11n was used as the baseline and replaced with several alternative structures. The corresponding results are presented in Table 7. The C2PSA baseline achieves 69.9% mAP50% and 35.9% mAP50-95. After replacing C2PSA with C2DA, the model obtains 69.8% mAP50% and 35.8% mAP50-95. C2BRA and C2CGA achieve mAP50 values of 70.1% and 70.2%, respectively, with corresponding mAP50-95 values of 35.9% and 35.8%. Under the same experimental settings, C2TSSA achieves 70.6% mAP50% and 36.2% mAP50-95, representing increases of 0.7 and 0.3 percentage points, respectively, compared with the C2PSA baseline. Among the compared modules, C2TSSA obtains the highest values for both evaluation metrics in this experiment. These results indicate that the C2TSSA-based configuration provides a modest improvement in detection performance under the current dataset and experimental conditions.

TABLE 7. Ablation experiments on different feature enhancement modules.

Furthermore, to confirm the efficacy of the suggested detection module, comparative evaluations involving diverse detection architectures were performed within the identical baseline structure. The outcomes are shown in Table 8. The standard detection head attained an mAP50 value of 69.9%, featuring 2.6 M parameters and a processing cost of 6.3 GFLOPs. EfficientHead minimized both parameter counts and operational loads to 2.3 M and 5.1 GFLOPs accordingly. Nevertheless, its mAP50 dropped to 69.4%, suggesting that overly streamlined designs might diminish feature interpretability. SEAMHead likewise lowered model intricacy but suffered a slight precision drop, obtaining an mAP50 of 69.6%. MultiSEAMHead matched the Baseline’s mAP50, both reaching 69.9%; yet, its parameter volume rose significantly to 4.5 M. This implies that the supplementary modules integrated into this architecture failed to provide significant performance enhancements. Conversely, the proposed detection module boosted mAP50% to 70.5% while preserving the identical parameter quantity and computational expense as the basic head. This represents gains of 0.6, 1.1, 0.9, and 0.6 percentage points over the Baseline, EfficientHead, SEAMHead, and MultiSEAMHead, respectively. Such findings demonstrate that the proposed detection module improves target distinctiveness without augmenting model complexity, consequently achieving a superior accuracy-efficiency balance.

TABLE 8. Ablation experiments on different detection heads.

3.4 Comparative experiments

To further verify the performance advantages of the GSDD-Net algorithm, a series of comparative experiments were conducted on the real-world gear defect dataset, GID dataset. The experimental configuration and training parameters used in these experiments were kept consistent with those described in Section 3.2.

Figure 7 compares the P-R curves of the baseline YOLO11n and GSDD-Net on the test set, with the results for YOLO11n shown on the left and those for GSDD-Net on the right. Both models maintained relatively high precision across most of the recall range, while the GSDD-Net curves enclosed a larger overall area. The overall mAP50 increased from 69.9% for YOLO11n to 73.0% for GSDD-Net. For the hp_cm, hp_cd, and kp classes, AP increased from 73.1%, 69.1%, and 67.6%–74.8%, 72.2%, and 71.9%, respectively. Among these classes, kp showed the largest numerical increase of 4.3 percentage points. Under the reported test conditions, GSDD-Net therefore achieved higher overall and class-specific AP values than the baseline model.

Figure 8 compares the normalized confusion matrices of the baseline YOLO11n and GSDD-Net on the test set. The diagonal entries represent the proportions of correctly detected and classified instances in each class, while the off-diagonal entries reflect misclassifications, false positives, or missed detections. For hp_cm, hp_cd, and kp, the diagonal values increased from 0.75, 0.69, and 0.65 for YOLO11n to 0.79, 0.76, and 0.71 for GSDD-Net, respectively. These increases indicate higher proportions of correct detections across all three defect classes. The displayed matrices also show lower normalized values for background-related false positives and missed detections with GSDD-Net. Overall, this comparison indicates improved detection performance on the test set under the evaluation settings used for Figure 8.

To evaluate the overall performance of the proposed approach, GSDD-Net was compared with several representative object detection models, and the quantitative results are presented in Table 9. Among the evaluated models, GSDD-Net achieved precision, recall, F1-score, mAP50, and mAP50–95 values of 74.5%, 67.7%, 70.8%, 73.0%, and 38.0%, respectively, showing competitive performance across the considered accuracy metrics. Compared with the baseline YOLO11n, the mAP50 and mAP50–95 increased by 3.1 and 2.1 percentage points, respectively, while the number of parameters decreased from 2.6 M to 1.9 M and GFLOPs decreased from 6.3 to 5.9; meanwhile, the inference speed increased from 93.7 to 105.7 FPS. These results indicate that GSDD-Net improves detection performance while maintaining relatively low model complexity and computational cost under the current experimental setting. Compared with the other lightweight detection models included in Table 9, different methods exhibited different trade-offs between detection accuracy and computational efficiency, with some models providing advantages in inference speed or computational cost and others achieving competitive accuracy with greater model complexity. Within this comparison, GSDD-Net achieved the highest mAP50 and mAP50–95 with 1.9 M parameters, suggesting a favorable balance among detection accuracy, parameter size, computational cost, and inference speed. Overall, the comparative results demonstrate that GSDD-Net provides competitive performance among the evaluated models and shows potential for lightweight object detection applications.

TABLE 9. Comparison of mainstream detection algorithms on the GID dataset.

Table 10 reports the category-wise AP50 and AP50-95 results of the compared models. The proposed GSDD-Net achieves superior or highly competitive performance on the hp_cm, hp_cd, and kp categories, indicating that its improvements are not limited to the overall detection score but also extend to category-specific recognition and localization. Under the AP50 metric, GSDD-Net obtains 72.2%, 71.9%, and 74.8% for kp, hp_cm, and hp_cd, respectively. These results represent improvements of 1.7, 3.1, and 4.3 percentage points over YOLO11n. Among the three categories, kp exhibits the largest improvement, suggesting that the proposed architecture provides more discriminative feature representations for this defect type. The advantages of GSDD-Net are further reflected by the stricter AP50-95 metric. Specifically, it achieves 40.6%, 42.3%, and 31.0% for hp_cm, hp_cd, and kp, respectively, corresponding to gains of 2.4, 1.4, and 2.4 percentage points over YOLO11n. Since AP50-95 evaluates localization at multiple IoU thresholds, these improvements indicate that GSDD-Net enhances not only defect recognition but also bounding-box localization accuracy and robustness.

TABLE 10. Comparison of AP50 and AP50-95 metrics for different models in class-wise detection tasks.

The performance variations among the compared models reflect their different abilities to handle inter-category differences in defect size, texture, and shape. In contrast, GSDD-Net improves both evaluation metrics for all three categories without a category-specific degradation in this experiment, demonstrating relatively consistent class-wise performance. Nevertheless, deep-learning results may fluctuate slightly because of random initialization, mini-batch ordering, stochastic data augmentation, and GPU-related nondeterminism. To reduce these effects, all compared models were trained under the same data split, training configuration, and hardware–software environment. Therefore, the results in Table 9 should be interpreted as a controlled comparison under identical experimental conditions. Overall, the category-wise results support the effectiveness of GSDD-Net in improving both detection and localization performance across different gear-defect categories.

Furthermore, to comprehensively assess the generalizability and robustness of the proposed GSDD-Net architecture, several benchmarking tests were executed using the publicly accessible gear defect dataset. The experimental setup and training hyperparameters were maintained identical to the specifications outlined in Section 3.2. Table 11 presents the comparison between GSDD-Net and several representative object detection models on this benchmark dataset. The results show that different lightweight YOLO-based models exhibit different trade-offs between detection accuracy and computational efficiency. For example, HyperYOLO-t achieves a recall of 97.0% and an mAP50–95 of 68.3%, with 3.2 M parameters and 9.6 GFLOPs. YOLOv12n obtains an mAP50 of 97.3% with 5.8 GFLOPs, while its mAP50-95 is 67.6%. RT-DETR-1 achieves an mAP50 of 97.0%, but requires 32.0 M parameters and 103.4 GFLOPs. Under the same evaluation setting, GSDD-Net achieves precision, recall, F1-score, mAP50, and mAP50–95 values of 95.6%, 97.9%, 96.7%, 97.8%, and 69.6%, respectively. Compared with the baseline YOLO11n, mAP50 and mAP50–95 increase by 0.7 and 2.1 percentage points, respectively, while the number of parameters decreases from 2.6 M to 1.9 M and GFLOPs decreases from 6.3 to 5.9. Compared with HyperYOLO-t, GSDD-Net achieves a 1.3 percentage-point improvement in mAP50–95 while using fewer parameters and lower computational cost. These results indicate that GSDD-Net maintains competitive detection performance and relatively low model complexity on the public gear defect dataset, with favorable performance under stricter IoU evaluation criteria. Overall, the additional benchmark results suggest that the proposed model can maintain a reasonable balance between detection accuracy and computational efficiency across the evaluated datasets, providing further evidence of its potential applicability to gear defect detection tasks.

TABLE 11. Comparative results of mainstream detection algorithms on the public gear defect dataset.

Figure 9 compares the P-R curves of YOLO11n and the proposed model, with the left panel corresponding to YOLO11n and the right panel corresponding to the proposed model. Based on the provided diagram, the comprehensive P–R curve for our developed framework is positioned nearer to the upper-right quadrant and encompasses a broader area compared to the YOLO11n baseline. The mAP50 metric across various classes rises from 97.1% to 97.8%, demonstrating that the enhanced architecture further boosts global recall while preserving superior detection accuracy. Regarding the classification-specific outcomes, the AP for the break category reaches 99.5% for both frameworks, implying that this group possesses distinct features and that the novel model sustains consistent detection efficiency. The AP for the lack category also stays at 99.5%, suggesting that the architectural refinements do not impair the identification results for this specific class. For the scratch category, which represents a more challenging target to detect, the AP climbs from 92.3% to 94.4%, equating to a 2.1 percentage point gain. This shows that the proposed framework can more efficiently identify subtle defect characteristics and refine the recognition of weak-texture objects. Collectively, while upholding high-precision identification for the break and lack categories, our model substantially elevates the detection efficiency for the scratch category, thus refining the total P-R curve and further improving its holistic detection capabilities.

In addition, to further evaluate the overall performance of GSDD-Net, several detection algorithms that achieved relatively competitive results on the GID dataset were selected for comparative experiments based on the results reported in Table 9, including YOLOv8n, YOLO-World, YOLO-Worldv2, HyperYOLO-t, and RT-DETR-l. Similarly, according to the results in Table 11, YOLO11n, YOLOv12n, YOLO-Worldv2, HyperYOLO-t, and RT-DETR-l were selected for comparison on the public gear defect dataset. The corresponding visual detection results are presented in Figures 10, 11, which show the performance of different detection models on the GID dataset and the public gear defect dataset, respectively. The visual results indicate that GSDD-Net is generally able to detect different types of gear surface defects and localize the corresponding defect regions on both datasets, while maintaining relatively few missed detections and false positives in the illustrated samples. Compared with the selected benchmark models, the bounding boxes predicted by GSDD-Net show relatively good agreement with the ground-truth annotations in several representative cases. In contrast, some benchmark models exhibit occasional incomplete detections, localization deviations, or classification errors, particularly for small defects, defects with weak texture characteristics, and defects occurring in relatively complex backgrounds. Overall, the visual comparisons suggest that GSDD-Net can effectively capture relevant features of gear surface defects and maintain relatively stable detection and localization performance across the two datasets. These qualitative observations are consistent with the quantitative results reported above and provide additional support for the applicability of GSDD-Net to gear surface defect detection.

3.5 Actual detection results

To validate the empirical effectiveness of GSDD-Net regarding gear surface flaw identification, various gear images featuring representative flaws, such as break, lack, scratch, kp, hp_cm, and hp_cd, were chosen for visual evaluation. The identification outcomes are displayed in Figure 12. The upper row presents the identification results of the reference framework YOLO11n, while the lower row displays those of GSDD-Net. In the selected examples, YOLO11n shows missed recognitions, erroneous identifications, and positioning errors, particularly for small-scale flaws and subtle defects in complex metallic-texture environments. GSDD-Net identifies more of the shown gear surface flaws, and several predicted boxes are more closely aligned with the annotated defect areas. This qualitative difference is consistent with the intended roles of ADown, CSP-EDLAN, C2TSSA, and LQEHead in retaining fine details, integrating multi-scale characteristics, incorporating contextual information, and using localization quality. The examples provide qualitative support for the reported precision and inference-speed results under the evaluated settings, without establishing performance for untested industrial conditions.

Furthermore, Grad-CAM heatmaps were employed to compare the feature responses of GSDD-Net and the baseline YOLO11n. As shown in Figure 13, the high-response regions of GSDD-Net are mainly concentrated on the defect areas and closely align with the actual defect locations, demonstrating its capability to capture discriminative local features of various defects, including break, lack, scratch, kp, hp_cm and hp_cd. In contrast, YOLO11n exhibits more scattered activations, with attention partially distributed over irrelevant background regions and weaker responses to some low-saliency defects. These results indicate that GSDD-Net can better distinguish defect-related features from complex background information and achieve more accurate feature localization. This improvement benefits from the complementary effects of the proposed components: ADown preserves contour and texture details during downsampling, CSP-EDLAN enhances feature propagation and multi-scale interaction, C2TSSA strengthens contextual feature representation while suppressing background interference, and LQEHead improves classification-localization coordination through localization quality estimation. Overall, the visualization results demonstrate that GSDD-Net achieves more effective defect-focused feature representation and robust localization under complex gear surface conditions.

4 Conclusion

This study introduces GSDD-Net, a specialized gear surface defect recognition framework for intricate industrial environments, designed to identify common flaws such as break, lack, scratch, kp, hp_cm, and hp_cd with high precision and speed. To tackle substantial size disparities in gear imperfections, loss of fine-scale details, the vulnerability of faintly defined defects to disturbances from complicated metallic-patterned substrates, and the challenge of achieving a perfect trade-off between recognition accuracy and computational speed, GSDD-Net enhances the YOLO structure across four dimensions: feature reduction, multi-resolution representation, global context integration, and localization-quality consciousness. Specifically, the ADown module is utilized to reformulate the feature compression procedure in a streamlined fashion, thus ensuring better maintenance of boundary outlines and texture details for delicate defects. The CSP-EDLAN module merges cross-stage partial links, multi-gradient path pooling, and the DualConv feature refinement technique to boost gradient flow, channel engagement, and multi-level semantic synthesis. The C2TSSA module constructs global context relationships with linear complexity utilizing the Token statistical self-attention strategy, strengthening the recognition of faintly visible defects and mitigating background noise triggered by metallic glints, lubricant residues, and non-uniform lighting. Additionally, LQEHead is developed to realize flexible synchronization between classification certainty and bounding-box positioning precision via localization quality assessment, thereby enhancing defect positioning precision and identification robustness.

The empirical findings indicate that, regarding the practical gear fault dataset, specifically the GID dataset, GSDD-Net attains Precision, Recall, F1-score, mAP50, and mAP50-95 metrics of 74.5%, 67.7%, 70.8%, 73.0%, and 38.0%, correspondingly. In comparison with the YOLO11n reference framework, GSDD-Net enhances mAP50 and mAP50-95 by 3.1 and 2.1 percentage points, respectively. Furthermore, the parameter count is minimized from 2.6 M to 1.9 M, GFLOPs drop from 6.3 to 5.9, and the processing velocity achieves 105.7 FPS. Compared with prevailing detection frameworks, GSDD-Net reaches the top mAP50 and mAP50-95 on the GID dataset while utilizing the least parameters. Within the open gear defect dataset, GSDD-Net obtains Precision, Recall, F1-score, mAP50, and mAP50-95 figures of 95.6%, 97.9%, 96.7%, 97.8%, and 69.6%, respectively, further validating its broad applicability and consistency. Such outcomes illustrate that GSDD-Net efficiently boosts the precision, resilience, and dependability of gear surface flaw identification in intricate industrial settings while preserving a slim architecture and rapid inference speed. Consequently, it offers a viable technical approach for developing a fully automated vision recognition framework for gear flaw identification.

Despite encouraging results in gear surface defect detection, GSDD-Net has several limitations that warrant further investigation. Future work should focus on: (1) expanding the dataset to cover more defect categories and sizes, varied lighting conditions, and complex industrial environments to improve robustness and generalizability; (2) exploring semi-supervised learning, weakly supervised learning, and synthetic defect generation to reduce the cost of acquiring high-resolution annotated data and address sample scarcity and class imbalance; (3) investigating model pruning, quantization, and edge deployment to reduce computational costs and inference latency on resource-constrained inspection hardware; and (4) exploring the integration of GSDD-Net with large language models to develop an intelligent agent for gear defect detection within an integrated visual inspection framework and assess its practical utility.

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: Viet Q. Vu, Thai Nguyen University of Technology, Vietnam

Reviewed by: Qian Chen, Zhejiang Normal University, China

Rui Shi, Anhui University of Science and Technology, China