Adaptive large-kernel convolutional network with gaussian-modulated spatial calibration for tyre quality assessment

1 Introduction

Tyre manufacturing quality directly impacts road safety, fuel efficiency, and vehicle longevity. Surface defects such as cracks, bulges, wear patterns, and foreign object intrusions can lead to catastrophic failures if undetected during production or maintenance [1,2]. According to industry statistics, tyre-related defects account for approximately 15% of all vehicle equipment failures, with sidewall cracks and tread separations being the predominant causes of high-speed blowouts. Traditional manual inspection methods are labor-intensive, subjective, and prone to operator fatigue, especially when examining the complex anisotropic multi-textured rubber layers present in both sidewall and tread regions [3]. A skilled inspector typically evaluates 200–300 tyres per hour under controlled lighting, but accuracy degrades significantly after 90 min of continuous work due to visual fatigue. Consequently, the development of automated, high-accuracy, and robust tyre defect detection systems has become an urgent priority for the automotive industry, with potential economic benefits estimated at $2.3 billion annually in reduced warranty claims and recall prevention.

In recent years, deep learning-based computer vision methods have demonstrated remarkable success in industrial defect detection tasks [4–6]. Convolutional Neural Networks (CNNs), including VGGNet [7], ResNet [5], DenseNet [8], and EfficientNet [9], have been widely adopted as backbone architectures for feature extraction. These networks learn hierarchical representations directly from raw pixel data, eliminating the need for hand-crafted feature engineering that previously limited the adaptability of machine vision systems. More recently, Vision Transformers (ViTs) [10] and their hybrid variants have introduced global self-attention mechanisms that capture long-range dependencies more effectively than pure convolutional operations. However, these advances come with significant computational costs and often require large-scale pre-training datasets, which are scarce in specialized industrial domains such as tyre manufacturing [6]. The quadratic complexity of self-attention with respect to image resolution makes ViTs particularly challenging to deploy on high-resolution industrial imagery, where pixel-level defect details must be preserved.

Several challenges persist in applying existing deep learning models to tyre defect classification.

• Limited receptive field: Standard CNNs with small fixed kernels (e.g., 3 × 3) struggle to capture global contextual information, which is essential for distinguishing between genuine defects and texture variations.
• Dynamic appearance variations: Tyre surface images exhibit substantial illumination differences, viewpoint changes, and rubber compound variations across manufacturing lines, demanding robust feature modulation mechanisms.
• Small sample sizes: Publicly available tyre defect datasets are limited in scale compared to general-purpose benchmarks like ImageNet, making it difficult to train deep models without overfitting.
• Inefficient architecture design: Existing large-kernel networks such as RepLKNet [11] and UniRepLKNet [12] improve receptive fields but incur heavy computational burdens, limiting real-time industrial deployment.

To address these challenges, we propose UniConvNet, a novel adaptive large-kernel convolutional network specifically designed for tyre quality assessment. The core contributions are summarized as follows.

• We introduce an OverLoCK ContMix Dynamic Convolution module that replaces traditional static receptive field attention (RFA) mechanisms. By computing token-wise affinity maps with respect to region centers, ContMix generates spatially varying dynamic kernels that encode global context while preserving strong local inductive biases.
• We propose a Gaussian Distribution Modulation module that explicitly models feature uncertainty under varying imaging conditions, followed by Spatial Feature Calibration to recalibrate channel-wise responses and suppress background interference.
• We design a Progressive Large Kernel Regulation strategy that progressively increases kernel sizes from 9 × 9 to 11 × 11 through a multi-branch feature aggregation structure. This approach balances fine-grained local detail with coarse global semantics without the computational overhead of monolithic large kernels.
• We develop a Multi-scale Feature Fusion mechanism that aligns and compresses features from three network stages via channel alignment, concatenation, and 1 × 1 convolution, enabling comprehensive multi-scale representation learning.
• Extensive experiments on the TyreNet dataset and a public Roboflow tyre defect dataset demonstrate that UniConvNet achieves 96.45% accuracy on TyreNet, surpassing ResNet-50 (90.12%), ConvNeXt-T (93.68%), and OverLoCK-T (94.89%) while maintaining lower computational complexity.

The remainder of this paper is organized as follows. Section 2 reviews related work on industrial defect detection, dynamic convolution, and large-kernel networks. Section 3 presents the proposed UniConvNet architecture in detail, including mathematical formulations and module designs. Section 4 describes datasets, implementation details, and experimental results, including ablation studies, comparisons with state-of-the-art methods, cross-dataset generalization, and visual analysis. Section 5 discusses the implications, limitations, and industrial deployment considerations. Section 6 concludes the paper.

2 Related work

2.1 Industrial defect detection methods

Industrial visual inspection has evolved from traditional machine vision approaches to modern deep learning systems. Early methods relied on hand-crafted features such as Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and wavelet transforms [13]. While effective for specific defect types, these approaches lack generalization across diverse materials and illumination conditions.

The advent of deep learning revolutionized defect detection. Wang et al. [14] proposed a Fully Convolutional Network (FCN) for tyre X-ray image segmentation, achieving pixel-wise defect localization. Li et al. [2] introduced TireNet, an end-to-end classification network using ResNet-50 and ResNet-101 backbones, reporting a detection omission rate of only 0.17% on a proprietary X-ray dataset. Saleh et al. [1] conducted a comprehensive transfer learning study across nine pre-trained models, with Xception achieving the best accuracy of 94.75% on their proprietary tyre dataset. Liu et al. [15] improved ShuffleNet for tyre crack detection, reaching 94.7% accuracy. However, these methods predominantly rely on standard fixed-kernel convolutions or generic transfer learning, without exploiting the potential of adaptive receptive fields for tyre-specific defect patterns.

Beyond tyre inspection, deep learning has been successfully applied to surface defect detection in steel manufacturing [16], fabric quality control [6], and printed circuit board (PCB) inspection [8]. These applications share common challenges with tyre defect detection, including small defect sizes, complex textured backgrounds, and real-time processing requirements. However, tyre surfaces present unique challenges due to their curved geometry, rubber-specific texture patterns, and the presence of embossed markings and tread grooves that can be mistaken for defects. These domain-specific characteristics motivate the need for specialized architectures rather than off-the-shelf backbones.

2.2 Dynamic convolution and attention mechanisms

Dynamic convolution aims to enhance feature representation by adapting kernel weights based on input content. CondConv [17] and Dynamic Convolution [18] introduced input-dependent filters that vary across channels. Involution [19] proposed spatially varying kernels generated per pixel, offering fine-grained adaptability at the cost of increased computation. Despite these advances, existing dynamic convolutions primarily operate within local neighborhoods and lack explicit global context modeling, which limits their effectiveness for defects that require holistic scene understanding.

The self-attention mechanism in Vision Transformers [10,20] provides an alternative approach to global modeling, computing pairwise interactions between all spatial tokens. While powerful, self-attention incurs quadratic complexity with respect to image resolution and sacrifices the translation equivariance and local inductive biases that make convolutions efficient for dense prediction tasks. Several hybrid architectures have attempted to combine the strengths of both paradigms, such as CoAtNet [21] and BoTNet [22], but these approaches remain computationally demanding for industrial deployment.

More recently, Lou and Yu [23] presented OverLoCK, a pure ConvNet backbone featuring Context-Mixing Dynamic Convolution (ContMix). ContMix calculates affinity values between each feature token and a set of region centers in a context feature map, then transforms these affinities into spatially varying dynamic convolution kernels. This mechanism enables long-range dependency modeling while preserving the inherent inductive bias of convolutions. OverLoCK-T achieved 84.2% top-1 accuracy on ImageNet-1K, surpassing ConvNeXt-B with only one-third of the FLOPs. We adapt and extend this ContMix mechanism within the tyre-specific architecture, integrating it with multi-scale channel splitting and Gaussian modulation to address the unique challenges of industrial tyre inspection.

Beyond industrial inspection, recent studies in scientific image analysis have similarly demonstrated the effectiveness of content-adaptive feature modeling under limited-data regimes, such as diffusion-based generative modeling with temporal dependency encoding and residual networks equipped with dynamic graph attention for spatially structured prediction tasks [24,25]. These works also exemplify reproducible research practices, including publicly released source code and documented data splits, which we follow in this study.

2.3 Large kernel convolutional networks

Expanding convolutional kernel sizes has emerged as an effective strategy for increasing receptive fields without abandoning convolutional inductive biases. ConvNeXt [26] demonstrated that 7 × 7 depthwise convolutions can rival Swin Transformer performance. RepLKNet [11] pushed kernel sizes to 31 × 31, revealing the promising performance of very large kernels. UniRepLKNet [12] further explored large kernels with dilated reparameterization, achieving competitive results on dense prediction tasks.

However, monolithic large kernels suffer from two drawbacks: (1) excessive computational cost, especially on high-resolution industrial images, and (2) limited adaptability to varying defect scales. The computational complexity of a direct K × K convolution scales as O ( K 2 ), making kernels beyond 11 × 11 prohibitively expensive for high-resolution inspection. Furthermore, static large kernels apply uniform weights across the entire feature map, regardless of local content, which is suboptimal for tyre images where fine cracks require localized small-kernel processing while bulges and structural deformations benefit from broader context.

The Progressive Large Kernel Regulation addresses these issues by decomposing large-kernel computation into multi-branch feature aggregation combined with a static 11 × 11 kernel, modulated by learned gating weights. This approach draws inspiration from structural reparameterization techniques [27], where multiple parallel branches with different characteristics are fused during inference. By progressively expanding the effective receptive field through stacked dynamic and static large-kernel stages, UniConvNet captures both micro-scale crack details and macro-scale structural anomalies within a unified framework.

2.4 Feature modulation and calibration

Feature modulation techniques enhance model robustness by dynamically adjusting feature responses. Squeeze-and-Excitation (SE) networks [28] recalibrate channel-wise features using global average pooling. CBAM [29] extended this with spatial attention. More recently, Gaussian-based modulation has been explored for uncertainty-aware feature enhancement in medical imaging [30]. We introduce Gaussian Distribution Modulation to explicitly model the statistical distribution of tyre surface features, enabling the network to adaptively attenuate noisy regions and amplify defect-sensitive activations.

3 Methodology

3.1 Problem definition

Given an input tyre surface image x ∈ R H × W × 3, the objective is to predict its defect class label y ∈ Y = { 1,2 , … , C }, where C denotes the number of defect categories. For the primary TyreNet dataset, C = 2 (good versus defective), while for the supplementary Roboflow dataset, C = 4 (normal, crack, bulge, wear).

The classification network f θ : R H × W × 3 → R C maps the input image to a probability distribution over classes via the softmax function as Equation 1:where θ represents all learnable parameters. During training, we minimize the cross-entropy loss between predicted probabilities and ground-truth labels.

p y = c ∣ x ; θ = exp f θ c x ∑ j = 1 C ⁡ exp f θ j x , (1)

3.2 Overall architecture overview

Figure 1 presents the overall architecture of UniConvNet. The network comprises six functional stages: (1) initial embedding and downsampling, (2) Stage-1 dual-branch processing with multi-scale channel split and OverLoCK ContMix, (3) Gaussian Distribution Modulation and Spatial Feature Calibration, (4) Stage-2 dual-branch processing with 9 × 9 dynamic kernels, (5) Progressive Large Kernel Regulation 9 × 9 → 11 × 11, and (6) Multi-scale Feature Fusion with task-specific classification head.

To make the technical contribution of UniConvNet explicit, we clarify which components are directly adopted from prior work, which are modified, and which are newly designed for tyre quality assessment. Directly adopted from OverLoCK [23]: the core ContMix principle, i.e., computing token-wise affinities between spatial queries and S × S region centers ( S = 7 ) and converting these affinities into spatially varying dynamic kernels through a learnable linear transformation with grouped matrix multiplication ( G = 4 ). Modified from the original design: (i) the ContMix branch is embedded within a dual-branch structure alongside a parallel multi-scale channel split branch, rather than serving as a standalone processing block; (ii) the dynamic kernel size follows a progressive schedule (3 × 3 in Stage-1 and 9 × 9 in Stage-2) instead of a fixed kernel configuration; and (iii) the dynamic branch output is fused with the static branch via batch normalization and a 1 × 1 projection to stabilize training under the small-sample conditions typical of industrial inspection. Newly designed for tyre quality assessment: the Gaussian Distribution Modulation and Spatial Feature Calibration modules for illumination-robust feature enhancement and background suppression, the Progressive Large Kernel Regulation with dilated multi-branch aggregation, the multi-scale feature fusion mechanism with channel alignment and 1 × 1 compression, and the auxiliary distillation loss that enforces prediction consistency between individual stages and the fused output.

3.2.1 Initial embedding

The input image x first passes through a 3 × 3 convolutional layer with stride one for initial feature embedding as Equation 2:followed by Batch Normalization (BN) and ReLU activation as Equation 3:

z 0 = Conv 3 × 3 x ; W e ∈ R H × W × C 0 , (2)

z 1 = ReLU BN z 0 . (3)

A downsampling layer with stride two reduces spatial dimensions to manage computational cost as Equation 4:

z 2 = Downsample z 1 ∈ R H 2 × W 2 × C 0 . (4)

3.2.2 Stage-1 dual-branch structure

The downsampled feature z 2 feeds into two parallel branches that capture complementary aspects of the tyre surface appearance. This dual-branch design is motivated by the observation that tyre defects manifest at vastly different scales: fine cracks may span only 10–20 pixels, while structural bulges and deformations occupy large continuous regions. A single branch with fixed receptive fields cannot optimally represent both extremes simultaneously.

Branch A - Multi-scale Channel Split: This branch splits the input feature along the channel dimension into two groups, processing each with different kernel sizes to establish a multi-scale feature pyramid at the single-stage level. Specifically, we employ a channel split ratio of γ : ( 1 − γ ), where γ = 0.5 in the implementation. As shown in Equations 5 and 6: where z 2 ( 1 ) captures point-wise channel interactions that model cross-channel correlations between different texture primitives, and z 2 ( 3 ) extracts local spatial patterns such as edge continuations and corner structures indicative of crack initiation. The outputs are concatenated along the channel dimension as shown in Equation 7:

z 2 1 = Conv 1 × 1 z 2 ; W 1 × 1 , (5)

z 2 3 = Conv 3 × 3 z 2 ; W 3 × 3 , (6)

f A 1 = Concat z 2 1 , z 2 3 ∈ R H 2 × W 2 × 2 C 0 . (7)

The channel splitting strategy differs from standard grouped convolutions by explicitly assigning different spatial scales to different channel groups, rather than merely partitioning computation. This design choice enables the network to learn scale-specific feature representations without requiring explicit multi-resolution input pyramids.

Branch B - OverLoCK ContMix Dynamic Kernel: This branch applies the Context-Mixing Dynamic Convolution from OverLoCK [23], which we adapt for tyre defect detection. Given the input feature z 2, we first transform it into query and key representations through learned linear projections. As shown in Equations 8 and 9: where Pool ( ⋅ ) denotes adaptive average pooling to S × S region centers (S = 7 in the implementation), and ∗ denotes convolution. The query representation preserves full spatial resolution, while the key representation aggregates semantically meaningful region centers that serve as context anchors.

Q = Reshape W q ∗ z 2 ∈ R C × H W , (8)

K = Reshape W k ∗ Pool z 2 ∈ R C × S 2 , (9)

The channels are divided into G groups ( G = 4 ), and affinity matrices are computed via grouped matrix multiplication as shown in Equation 10:

A g = Q g ⊤ K g ∈ R H W × S 2 , g = 1 , … , G . (10)

Each row of A g holds the affinity values between the i-th spatial token and all S 2 region centers, effectively encoding the relationship between local appearance and global semantic context. A learnable linear transformation converts affinities into dynamic kernels as shown in Equation 11:where k is the fixed spatial kernel size (e.g., k = 3 for Stage-1). The dynamic convolution output is computed via standard sliding-window convolution with token-specific weights as shown in Equation 12:

K dyn i = W dyn ⋅ A g i , : ∈ R k × k × C in × C out , (11)

f B 1 i = ∑ m , n K dyn i m , n ⋅ z 2 i + m , i + n . (12)

Branch Fusion: The two branch outputs are fused via element-wise summation followed by a 1 × 1 projection that restores the channel dimension as shown in Equation 13:

f 1 = Conv 1 × 1 BN f A 1 + f B 1 . (13)

The residual-like summation before projection encourages gradient flow and enables the network to learn incremental improvements from each branch without catastrophic interference. Batch normalization stabilizes training by normalizing the combined activations, which is important because the dynamic branch produces features with different statistical properties than the static multi-scale branch.

3.2.3 Feature enhancement: gaussian modulation and spatial calibration

Following the dual-branch processing, the Stage-1 output f ( 1 ) undergoes feature enhancement through two sequential modules designed to improve robustness against imaging variations and background clutter.

Gaussian Distribution Modulation: Industrial tyre inspection systems operate under variable lighting conditions, ranging from bright fluorescent illumination in modern factories to uneven natural lighting in outdoor service stations. These variations cause significant shifts in the feature distributions learned by convolutional networks, potentially masking subtle defect signatures or amplifying irrelevant texture patterns. To address this challenge, we model the spatial feature distribution as a Gaussian and apply learned modulation.

Given feature map f ( 1 ), we compute channel-wise mean μ and variance σ 2 across all spatial locations. As shown in Equations 14 and 15:

μ c = 1 H W ∑ i = 1 H ∑ j = 1 W f c , i , j 1 , (14)

σ c 2 = 1 H W ∑ i = 1 H ∑ j = 1 W f c , i , j 1 − μ c 2 . (15)

These statistics provide a compact representation of the global feature distribution for each channel. A learnable modulation gate g gauss generates channel-wise scaling factors based on these statistics as shown in Equation 16:where σ ( ⋅ ) is the sigmoid function, and W g ∈ R C × 2 C projects the concatenated statistics to modulation weights. The modulated feature is provided in Equation 17:where ⊙ denotes channel-wise multiplication. The additive formulation ( 1 + g ) ensures that the base features are preserved when the gate is near zero, while allowing amplification or attenuation based on the learned distribution. This is particularly effective for tyre images where bright specular reflections (high variance regions) should be attenuated and low-contrast crack regions (anomalous mean values) should be amplified.

g gauss = σ W g ⋅ μ , σ 2 + b g , (16)

f mod = f 1 ⊙ 1 + g gauss , (17)

Spatial Feature Calibration: While Gaussian modulation operates at the channel level, spatial calibration addresses background suppression and defect localization. Using both average and max pooling along the channel dimension captures complementary aspects of spatial importance. As shown in Equations 18 and 19: where s avg indicates regions with consistently high activation across channels (likely structural features), and s max highlights regions with peak activation in at least one channel (potentially discriminative defect signatures).

s avg = 1 C ∑ c = 1 C f mod , c , (18)

s max = max c f mod , c , (19)

The spatial attention map is computed by concatenating these two descriptors and processing through a 7 × 7 convolution followed by sigmoid activation as shown in Equation 20:The 7 × 7 kernel provides sufficient context to distinguish coherent defect regions from isolated noise pixels. The calibrated feature output as shown in Equation 21:This calibration effectively suppresses uniform rubber background regions while preserving or enhancing defect boundaries, creating a cleaner feature representation for downstream processing.

M spa = σ Conv 7 × 7 s avg ; s max . (20)

f 2 = f mod ⊙ M spa . (21)

3.2.4 Stage-2 dual-branch structure

Stage-2 follows an analogous dual-branch design but operates at a reduced spatial resolution with larger 9 × 9 dynamic kernels in the OverLoCK ContMix branch. The multi-scale channel split branch now employs 1 × 1 and 5 × 5 convolutions to capture broader context. As shown in Equations 22–24:

f A 2 = Concat Conv 1 × 1 f 2 , Conv 5 × 5 f 2 , (22)

f B 2 = OverLoCK - ContMix 9 × 9 f 2 , (23)

f 3 = Conv 1 × 1 BN f A 2 + f B 2 . (24)

3.2.5 Progressive large kernel regulation

The Progressive Large Kernel Regulation (PLKR) module further expands the receptive field to 11 × 11 through a branched decomposition strategy that avoids the computational explosion of direct large-kernel convolution.

Branch A - Multi-branch Feature Aggregation: We decompose the 11 × 11 kernel into multiple parallel branches with different dilation rates. As shown in Equations 25–28: where α k are learnable aggregation weights satisfying ∑ k α k = 1.

f agg 1 = Conv 3 × 3 , d = 1 f 3 , (25)

f agg 2 = Conv 3 × 3 , d = 2 f 3 , (26)

f agg 3 = Conv 3 × 3 , d = 3 f 3 , (27)

f A 3 = ∑ k = 1 3 α k ⋅ f agg k , (28)

Branch B - 11 × 11 Static Large Kernel: A direct depthwise 11 × 11 convolution captures coarse global patterns as shown in Equation 29:

f B 3 = DW - Conv 11 × 11 f 3 . (29)

Branch Fusion: The fused output combines both branches with residual connection as shown in Equation 30:where f ( 4 ) serves as the Main Out for downstream fusion.

f 4 = ReLU BN Conv 1 × 1 f A 3 + f B 3 + f 3 , (30)

3.2.6 Multi-scale feature fusion

To leverage complementary information from different processing stages, we fuse features from Stage-1 (f ( 1 ), Feature Out), Stage-2 (f ( 3 ), Feature Out), and the PLKR module (f ( 4 ), Main Out).

Channel Alignment: Each feature is first projected to a common channel dimension C fuse. As shown in Equations 31–33:

f ̃ 1 = Conv 1 × 1 f 1 ; C fuse , (31)

f ̃ 3 = Conv 1 × 1 f 3 ; C fuse , (32)

f ̃ 4 = Conv 1 × 1 f 4 ; C fuse . (33)

Concatenation and Compression: The aligned features are resized to the spatial resolution of the PLKR output ( H ′ × W ′ ) via bilinear interpolation (with align-corners disabled), concatenated, and compressed. As shown in Equations 34, 35:

f cat = Concat f ̃ 1 , f ̃ 3 , f ̃ 4 ∈ R H ′ × W ′ × 3 C fuse , (34)

f fused = Conv 1 × 1 f cat ; C fuse . (35)

The fused feature f fused is fed to the classification head, consisting of global average pooling and a fully-connected layer with softmax output.

3.3 Complexity analysis

The computational complexity of each UniConvNet stage is shown in Table 1. The total parameter count is 31.4 M, and the total FLOPs for a 224 × 224 input is 4.2 G. The progressive kernel strategy reduces FLOPs by 24% compared to using a single static 11 × 11 kernel throughout the network (5.5 G estimated), while the dynamic convolution adds only marginal overhead (0.3 G) relative to its significant accuracy improvement.

TABLE 1. Computational complexity of UniConvNet stages. Input resolution: 224 × 224 .

The bold values represent the optimal performance in the corresponding column.

3.4 Loss function design

We employ the standard cross-entropy loss for the classification task as shown in Equation 36:where N is the batch size and 1 { ⋅ } is the indicator function.

L CE = − 1 N ∑ i = 1 N ∑ c = 1 C 1 y i = c log ⁡ p y i = c ∣ x i ; θ , (36)

Additionally, we introduce an auxiliary distillation loss to encourage consistency between the multi-scale fusion output and individual stage outputs as shown in Equation 37:where p s and p fused are softmax probabilities from stage-s and the fusion output, respectively, and KL denotes the Kullback-Leibler divergence.

L aux = ∑ s ∈ 1,3,4 KL p s ‖ p fused , (37)

The total training loss is provided in Equation 38:where λ = 0.1 is determined via grid search on the validation set over the candidate range λ ∈ { 0.01 , 0.05 , 0.1 , 0.5 , 1.0 }. The value λ = 0.1 yielded the highest validation accuracy; smaller values rendered the auxiliary objective ineffective, whereas larger values over-weighted the distillation term and degraded the primary classification performance.

L = L CE + λ L aux , (38)

4 Experiments

4.1 Datasets

4.1.1 TyreNet dataset

The TyreNet dataset [31] is a publicly available high-quality annotated dataset for tyre defect classification. It comprises 1,698 tyre images collected from six car and bike service stations and two tyre showrooms, ensuring diversity in imaging environments and lighting conditions. Images were captured using an Android smartphone with a 48 MP triple rear camera.

The dataset is organized into two classes: 866 defective tyre images showcasing cracks, wear patterns, and other anomalies; and 832 good (non-defective) tyre images. All images were annotated by domain experts. We adopt a stratified 7:2:1 split for training, validation, and testing with a fixed random seed (seed = 42), resulting in 1,188 training images, 340 validation images, and 170 test images. Stratification preserves the nearly balanced class distribution (approximately 51% defective versus 49% good) across all three subsets. The exact split files (image lists for each subset) will be released together with the source code upon publication to enable exact replication of the reported results.

4.1.2 Roboflow tyre defect dataset

To validate cross-dataset generalization, the public “Tyre defect detection” dataset from Roboflow Universe [32] is used. It contains 3,069 images across multiple defect categories including cracks, bulges, and normal tyres. The dataset includes train/validation/test splits (87%/8%/4%) with preprocessing including auto-orientation, resizing to 640 × 640, grayscale conversion, and augmentation (horizontal/vertical flips, rotation ± 15 °, saturation/brightness/exposure adjustments). It remaps the annotations into four classes for consistent evaluation: Normal, Crack, Bulge, and Wear.

4.1.3 Data preprocessing and augmentation

All images are preprocessed using the following pipeline.

• 1. Resizing: Images are resized to 224 × 224 for TyreNet and 640 × 640 for the Roboflow dataset.
• 2. Normalization: Pixel values are normalized to [ 0,1 ] and standardized with ImageNet statistics μ = [ 0.485 , 0.456 , 0.406 ], σ = [ 0.229 , 0.224 , 0.225 ].
• 3. Augmentation: Training images undergo random horizontal flipping, rotation ( ± 10 ° ), color jitter (brightness 0.2, contrast 0.2, saturation 0.1), and random cropping.
• 4. Class balancing: This study applies weighted random sampling to address the slight class imbalance in TyreNet (866 vs. 832).

4.2 Implementation details

All experiments are conducted on a workstation equipped with an NVIDIA RTX 4090 GPU (24 GB VRAM), an Intel Core i9-13900K CPU, and 64 GB RAM. The software stack comprises Python 3.10, PyTorch 2.1, CUDA 12.1, and torchvision 0.16.

Hyperparameters: The proposed UniConvNet is trained for 100 epochs using the AdamW optimizer with an initial learning rate of 1 × 1 0 − 4, weight decay 5 × 1 0 − 5, and cosine annealing schedule with five-epoch warm-up. The batch size is 32 for TyreNet and 16 for the higher-resolution Roboflow dataset. Label smoothing ( ϵ = 0.1 ) is applied to improve generalization. Early stopping with patience 15 is used based on validation loss.

Evaluation metrics: This study reports classification Accuracy, Precision, Recall, and F1-score. For multi-class evaluation on Roboflow, the macro-averaged metrics is used. Statistical significance is assessed via paired t-tests with p < 0.05.

4.3 Ablation studies

The comprehensive ablation studies are conducted on the TyreNet validation set to verify the contribution of each proposed module.

4.3.1 Component-wise ablation

Table 2 presents the incremental contribution of each architectural component, starting from a ResNet-50 baseline.

TABLE 2. Component-wise ablation study on TyreNet validation set. Baseline: ResNet-50. Each row adds one component to the previous configuration.

The bold values represent the optimal performance in the corresponding column.

Each component contributes positively to the final performance. Notably, the OverLoCK ContMix module provides the largest single gain (+1.17% accuracy), validating the effectiveness of dynamic context-mixing for tyre defect patterns. The Gaussian Modulation module contributes +0.87% accuracy, demonstrating its value in handling the illumination variations present in the TyreNet dataset, which was collected from multiple service stations with diverse lighting setups. Spatial Feature Calibration adds another +0.66%, primarily by suppressing background rubber texture that can be confused with fine cracks. The complete UniConvNet achieves 96.45% accuracy with only 31.4 M parameters and 4.2 GFLOPs, demonstrating high parameter efficiency. The relatively modest parameter increase from the baseline (25.6 M–31.4 M, +22.7%) is justified by the substantial accuracy improvement of +5.2% absolute.

It is worth clarifying why adding the multi-scale fusion module increases the parameter count (30.1 M → 31.4 M) while reducing FLOPs from 4.8 G to 4.2 G. Without fusion, the classification head must consume the full-channel PLKR output and therefore requires an additional channel-expansion convolution operating at 56 × 56 resolution to bridge the backbone features and the classifier. With aligned multi-scale fusion, the classifier instead consumes the compressed C fuse-dimensional fused feature, so this expensive full-channel spatial convolution is no longer needed. The three channel-alignment 1 × 1 convolutions contribute additional stored parameters but relatively few FLOPs, because 1 × 1 convolutions are computationally inexpensive at reduced spatial resolution. The net effect is thus slightly more parameters but lower computational overhead.

The progressive kernel strategy clearly achieves the optimal accuracy-efficiency trade-off, with the 9 × 9 → 11 × 11 configuration outperforming all fixed-kernel alternatives. Figure 2 provides visual comparisons of the component-wise and kernel size ablations.

4.3.2 Progressive kernel size ablation

The impact of kernel size progression is provided in Table 3. Directly applying large static kernels (11 × 11 or 13 × 13) yields diminishing returns and increased computational cost. This phenomenon occurs because monolithic large kernels apply uniform weights across diverse spatial locations, failing to adapt to the varying scale requirements of different tyre regions. The progressive strategy (9 × 9 dynamic + 11 × 11 static) achieves the best accuracy-efficiency trade-off by combining the adaptability of dynamic medium kernels with the global context aggregation of static large kernels.

TABLE 3. Progressive kernel size ablation on TyreNet validation set.

The bold values represent the optimal performance in the corresponding column.

The 13 × 13 kernel performs worse than 11 × 11 (94.35% vs. 94.89%), suggesting that excessive receptive fields introduce irrelevant background noise and dilute fine defect signals. This observation supports the design choice of capping the maximum static kernel at 11 × 11 and relying on the ContMix dynamic mechanism to extend effective receptive fields adaptively. All latency values reported in Table 3 are measured on the RTX 4090 with batch size one and represent the pure model forward-pass time, averaged over 1,000 runs after 100 warm-up iterations; data preprocessing (resizing and normalization) is excluded and adds approximately 2.1 m per image when measured end-to-end. The latency measurements further confirm the practicality of the proposed approach: 11.5 m per image is 28% faster than the 11 × 11-only configuration (16.5 m) despite achieving +1.56% higher accuracy.

4.3.3 Multi-scale fusion strategy

Different feature fusion strategies are compared, as shown in Table 4. Fusing all three stage outputs with channel alignment and 1 × 1 compression yields optimal performance.

TABLE 4. Ablation on multi-scale feature fusion strategies.

The bold values represent the optimal performance in the corresponding column.

4.4 Comparison with state-of-the-art methods

All compared methods are trained from scratch under identical experimental conditions (same data splits, augmentation, and training protocol) to ensure fair comparison. Table 5 compares UniConvNet against established backbones and recent architectures on the TyreNet test set.

TABLE 5. Comparison with state-of-the-art methods on the TyreNet test set. The best results are in bold, and the second-best are underlined.

The bold values represent the optimal performance in the corresponding column.

UniConvNet outperforms all competitors, achieving a 1.56% absolute improvement over OverLoCK-T and a 6.33% improvement over ResNet-50. Importantly, the proposed model maintains the lowest FLOPs among top-performing methods (4.2 G vs. 5.5 G for OverLoCK-T), demonstrating superior efficiency. The parameter count (31.4 M) is comparable to ConvNeXt-T and UniRepLKNet-T, indicating that the performance gains stem from architectural innovation rather than increased capacity. Figure 3 provides visual comparisons of accuracy and the accuracy-efficiency trade-off.

4.5 Cross-dataset generalization

To assess model robustness, the proposed UniConvNet is evaluated on the Roboflow tyre defect dataset without fine-tuning (zero-shot) and with 10-epoch fine-tuning. Table 6 reports macro-averaged metrics.

TABLE 6. Cross-dataset generalization on the Roboflow tyre defect dataset (4-class).

UniConvNet achieves strong zero-shot generalization (76.85%), significantly outperforming OverLoCK-T (72.45%) and ResNet-50 (62.35%). After brief fine-tuning, the proposed model reaches 94.12% accuracy, confirming that the learned representations transfer effectively across imaging conditions and defect taxonomies. The gap between zero-shot and fine-tuned performance is smallest for UniConvNet (19.27% vs. 22.0% for OverLoCK-T and 25.0% for ResNet-50), indicating more transferable feature learning.

UniConvNet converges faster and achieves lower validation loss than ResNet-50, benefiting from the auxiliary distillation loss and progressive feature refinement. Figure 4 shows cross-dataset comparison bar charts and training convergence curves.

4.6 Visual analysis

4.6.1 Grad-CAM visualization

UniConvNet accurately localizes crack regions and suppresses background textures, confirming that the OverLoCK ContMix dynamic kernels and Gaussian modulation successfully focus on discriminative defect features. For the defective sample shown in Figure 5, the Grad-CAM [34] heatmap intensity peaks precisely along the crack trajectory, with minimal activation in the surrounding rubber texture and tyre shoulder regions. This localization accuracy is particularly important for industrial applications where human operators must verify automated decisions: the Grad-CAM output provides an intuitive visual explanation that aligns with human perceptual judgment.

For good tyre samples, the heatmap exhibits diffuse low-intensity activation across the tread pattern, without spurious peaks that would indicate false defect detection. The Spatial Feature Calibration module plays a crucial role here by suppressing high-contrast but irrelevant textures such as tread grooves and manufacturing markings. Compared to ResNet-50, which often produces scattered activation patterns on good tyres due to its fixed-kernel limitations, UniConvNet demonstrates more focused and semantically meaningful attention maps.

4.6.2 Confusion matrix analysis

On TyreNet, only six out of 170 test samples are misclassified (3 false positives, three false negatives), yielding a balanced error rate of 3.53%. The symmetric error pattern (equal false positives and false negatives) indicates that the model does not exhibit class bias, which is important for industrial deployment where both missed defects and false alarms carry significant costs. Figure 6 shows confusion matrices for TyreNet (binary) and Roboflow (4-class).

On Roboflow, the majority of errors occur between visually similar classes (crack vs. wear), which is expected given their shared elongated appearance patterns. The Normal class achieves the highest precision (97.8%), while Bulge has the lowest recall (94.6%) due to the smaller sample size in the dataset. These patterns suggest that additional bulge-specific augmentation or class-weighted loss could further improve multi-class performance.

4.6.3 Error case analysis

This study manually inspects the misclassified samples to understand failure modes and guide future improvements. The primary failure modes are: (1) severe motion blur obscuring defect boundaries, typically occurring when images are captured during tyre rotation or with slow shutter speeds; (2) shallow scratches that are visually ambiguous even to human experts, often requiring tactile inspection for definitive classification; and (3) strong specular reflections creating false texture patterns that mimic crack edges, particularly problematic on glossy new tyres under direct lighting.

Interestingly, some good tyres with severe tread wear patterns are misclassified as defective, suggesting that the model has learned to associate wear with structural compromise even when no explicit crack or bulge is present. This behavior, while technically an error according to the ground-truth labels, may actually represent a desirable safety bias for industrial deployment, where conservative classification of marginal tyres is preferable to missed defects. These cases suggest that incorporating multi-frame temporal information or active illumination control could further improve robustness against challenging imaging conditions.

5 Discussion

The experimental results demonstrate that UniConvNet establishes a new state-of-the-art for tyre defect classification, achieving 96.45% accuracy on TyreNet. Several factors contribute to this performance:

Dynamic context modeling: The OverLoCK ContMix module enables each spatial location to adapt its receptive field based on semantic context, which is particularly beneficial for tyre images where defects vary dramatically in scale (fine cracks vs. large bulges). The affinity-based kernel generation ensures that dynamic weights carry global contextual information, allowing a 3 × 3 or 9 × 9 physical kernel to effectively model relationships across the entire image. This is especially important for distinguishing genuine sidewall cracks from innocuous molding seams, which requires understanding both local edge geometry and global structural continuity.

Progressive kernel expansion: Rather than abruptly introducing large kernels, the progressive strategy 3 × 3 → 5 × 5 → 9 × 9 → 11 × 11 allows the network to first establish local detail representations before aggregating global context, preventing information loss from premature spatial downsampling. The ablation study in Table 3 confirms that monolithic 11 × 11 kernels underperform compared to the progressive decomposition, validating the hypothesis that gradual expansion preserves fine-grained features while progressively incorporating coarse semantics.

Statistical robustness: Gaussian Distribution Modulation explicitly models feature uncertainty, providing built-in resilience against the illumination variations common in factory-floor imaging. By modulating channel activations based on learned statistical parameters (mean and variance), the network effectively normalizes feature responses across diverse imaging conditions without requiring explicit domain adaptation. The Spatial Feature Calibration further refines this by suppressing irrelevant background regions, concentrating computational resources on discriminative defect regions. Beyond illumination variation, physical tyre contaminants encountered in real-world service stations–such as heavy dirt, mud, or oil residues–can distort low-level feature distributions in a similar manner, shifting local intensity means and inflating variance in contaminated regions. Because the Gaussian modulation normalizes channel responses based on global statistics and the spatial calibration suppresses spatially incoherent activations, the model inherits partial robustness to such contamination-induced distribution shifts. Nevertheless, severe occluding contaminants that fully cover the defect region remain challenging, as reflected in our error case analysis, and pre-inspection cleaning or contamination-aware augmentation is advisable for deployment in uncontrolled service environments.

Multi-scale information integration: The feature fusion mechanism combines complementary representations from three processing stages. Early-stage features ( f ( 1 ) ) preserve fine spatial details critical for crack detection, mid-stage features ( f ( 3 ) ) capture texture patterns distinguishing defect types, and late-stage features ( f ( 4 ) ) encode holistic structural context. The channel alignment and 1 × 1 compression strategy prevents the parameter explosion that would result from naive concatenation, maintaining computational efficiency while enabling rich multi-scale representation learning.

Advantages: (1) Superior accuracy with competitive efficiency: UniConvNet outperforms all compared methods while using fewer FLOPs than OverLoCK-T. The accuracy improvement of +1.56% over OverLoCK-T and +6.33% over ResNet-50 is achieved with 31.4 M parameters, which is comparable to standard backbone networks. (2) Strong generalization: cross-dataset validation confirms that learned features are not overfitted to TyreNet’s imaging conditions. The zero-shot performance of 76.85% on the visually distinct Roboflow dataset demonstrates robust feature learning. (3) Interpretability: Grad-CAM visualizations reveal clear defect localization, facilitating human-in-the-loop quality assurance and regulatory compliance in safety-critical automotive applications. (4) Modular design: each component (ContMix, Gaussian modulation, progressive kernels) can be independently adopted in other industrial vision systems, enhancing the broader applicability of the contributions.

Limitations: (1) The current architecture is designed for image-level classification; extending to pixel-level defect segmentation would require additional decoder structures. (2) Real-time inference speed (11.5 m forward-pass latency per image on RTX 4090) is acceptable for offline inspection but may require further optimization for high-speed production lines exceeding 60 FPS. Quantization and TensorRT deployment could address this gap. (3) The model assumes relatively fixed camera viewpoints; significant perspective distortion or extreme close-up angles may degrade performance without viewpoint augmentation during training. (4) While effective for binary and coarse multi-class classification, fine-grained categorization of defect severity levels may require additional annotated data and hierarchical loss functions (5) Robustness to heavy physical contaminants (e.g., mud or dirt fully occluding the defect area). is limited; such cases may require pre-inspection cleaning or contamination-specific augmentation.

UniConvNet’s parameter count (31.4 M) and computational cost (4.2 GFLOPs) are compatible with edge deployment on modern industrial AI accelerators such as the NVIDIA Jetson AGX Orin or Intel Movidius VPU. The 11.5 m forward-pass latency on an RTX 4090 is estimated to translate to approximately 40–50 m on an NVIDIA Jetson AGX Orin (64 GB) with FP16 inference, based on the published relative throughput of the two platforms; this is sufficient for conveyor-belt inspection stations with moderate throughput (up to 20 tyres per minute). For higher-speed lines, quantization-aware training (INT8) and TensorRT optimization are recommended, which could reduce latency by 50%–60% with <0.5% accuracy degradation based on literature benchmarks [35].

From a system integration perspective, UniConvNet can be deployed as a drop-in replacement for existing ResNet-50 or EfficientNet backbones in tyre inspection pipelines, requiring only retraining on the target dataset. The modular architecture facilitates incremental upgrades: manufacturers can first deploy the multi-scale channel split and branch fusion components for immediate accuracy gains, then progressively add the OverLoCK ContMix and Gaussian modulation modules as computational budgets allow. This staged deployment strategy reduces the barrier to adoption for factories with legacy inspection infrastructure.

The interpretability provided by Grad-CAM visualizations also addresses an important practical concern in industrial AI deployment: quality engineers and regulators often require explanations for automated decisions. By generating attention maps that clearly localize defect regions, UniConvNet supports human-in-the-loop validation workflows and simplifies compliance with emerging AI transparency standards in automotive manufacturing.

Future research directions include: (1) extending UniConvNet to simultaneous defect detection and segmentation via a unified encoder-decoder architecture, enabling pixel-level defect boundary delineation for precise metrology; (2) incorporating semi-supervised learning to leverage abundant unlabeled tyre images from production lines, potentially reducing annotation costs by 60%–80% while maintaining accuracy; (3) investigating neural architecture search (NAS) to automatically optimize the progressive kernel schedule for specific defect taxonomies and imaging conditions; (4) developing lightweight student models through knowledge distillation from UniConvNet to teacher networks such as MobileNetV3 or EfficientNet-EdgeTPU, targeting sub-10 m inference on embedded devices; and (5) exploring active learning strategies that selectively request human annotation for the most informative samples, further reducing the annotation burden in production environments.

Additionally, we plan to investigate the integration of multi-spectral imaging (near-infrared and ultraviolet) with UniConvNet to detect subsurface defects invisible in standard RGB imagery. The dynamic convolution mechanism is particularly well-suited for multi-spectral fusion, as the affinity-based kernel generation can learn to emphasize spectral bands most informative for specific defect types. Finally, temporal consistency modeling across video sequences of rotating tyres could further improve detection reliability by exploiting the periodicity of tyre surface appearance.

6 Conclusion

This paper presented UniConvNet, an adaptive large-kernel convolutional network for automated tyre quality assessment. By integrating OverLoCK ContMix dynamic convolution, Gaussian Distribution Modulation, Spatial Feature Calibration, and Progressive Large Kernel Regulation, the architecture effectively addresses the dual challenges of fine-grained defect detail extraction and long-range contextual dependency modeling that have limited previous approaches in industrial tyre inspection.

The comprehensive experimental evaluation demonstrates the practical value of each architectural innovation. The OverLoCK ContMix module provides the largest individual accuracy gain (+1.17%) by enabling dynamic receptive field adaptation, allowing the network to simultaneously resolve fine cracks and contextualize large structural deformations. Gaussian Distribution Modulation contributes +0.87% by normalizing feature responses across diverse imaging conditions, while Spatial Feature Calibration suppresses background clutter that previously confused standard CNN backbones. The Progressive Large Kernel Regulation strategy achieves the optimal accuracy-efficiency trade-off, outperforming monolithic large-kernel configurations while using 24% fewer FLOPs. Finally, multi-scale feature fusion integrates complementary representations from three processing stages, preserving fine spatial details without excessive parameter growth.

From an industrial perspective, UniConvNet offers a practical deployment pathway for tyre manufacturers seeking to automate quality inspection. The modular architecture enables staged upgrades of existing inspection pipelines, while the computational efficiency (11.5 m forward-pass inference on a desktop GPU, estimated 40–50 m on a Jetson AGX Orin-class edge accelerator) supports real-time conveyor-belt inspection at moderate throughput. Future work will explore pixel-level defect segmentation, semi-supervised learning for annotation efficiency, and ultra-lightweight student models for high-speed production lines. The architectural principles introduced in this work can extend beyond tyre inspection to broader industrial vision applications where objects exhibit scale-varying defects under variable imaging conditions.

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: Yuxuan Ren, Fudan University, China

Reviewed by: Lin Yuan, Qilu University of Technology, China

Zhaolu Liu, Imperial College London, United Kingdom