High-Precision Fabric Defect Detection via Adaptive Shape Convolutions and Large Kernel Spatial Modeling

TL;DR

This work tackles the challenge of high-precision, real-time fabric defect detection by extending YOLOv8s with two novel modules: Adaptive Shape Convolution Module (ASCM) and Large Kernel Shift Convolution Module (LKSCM). ASCM enhances feature fusion in the Neck by dynamically adjusting convolutional shapes, while LKSCM emulates large-kernel behavior in the Backbone through a shift-convolution framework, yielding improved spatial modeling with lower compute. On the Tianchi fabric dataset, Fab-ASLKS achieves a $mAP@50$ of $60.3\%$, about a $4.9\%$ relative improvement over the Baseline, and surpasses other state-of-the-art methods while maintaining a modest parameter count ($10.5$M). The results demonstrate that synergistic integration of ASC-based adaptive receptive fields and shift-based large-kernel approximations yields robust, real-time defect detection across diverse and complex textile patterns, with particular gains for large and long defects. This approach holds practical significance for industrial quality control, enabling faster and more accurate automated inspection on production lines.

Abstract

Detecting fabric defects in the textile industry remains a challenging task due to the diverse and complex nature of defect patterns. Traditional methods often suffer from slow inference speeds, limited accuracy, and inadequate recognition rates, particularly in scenarios involving intricate or subtle defects. To overcome these limitations, we introduce Fab-ASLKS, an advanced fabric defect detection framework built upon the YOLOv8s architecture. Fab-ASLKS incorporates two key modules: (1) the Adaptive Shape Convolution Module (ASCM), which leverages adaptive shape convolution within the Neck to enhance feature fusion and improve efficiency by extending the capabilities of the standard C2f structure, and (2) the Large Kernel Shift Convolution Module (LKSCM), designed to emulate large kernel effects within the Backbone, enabling superior spatial information extraction. These modules collaboratively optimize feature extraction and information integration across the network. Extensive experiments conducted on the Tianchi fabric defect detection dataset demonstrate that Fab-ASLKS achieves a 5% improvement in mAP@50 over the baseline, showcasing its capability to deliver high precision and efficiency.

Figures (9)

• Figure 1: Different structures employed in fabric defect detection. (a) Traditional convolution for extracting local features. (b) Models based on transformers or large kernel convolutions extract global features. (c) Fab-ASLKS. We employ the original Conv-BN-SiLU (CBS) module in the YOLOv8 model to extract local features and large kernel shift convolution (LKSC) to capture global features. Meanwhile, the Adaptive Shape Convolution Module (ASCM) enhances the fourth C2f module in the Neck (PANet), enabling the dynamic adjustment of convolutional kernels and expanding the receptive field. These enhancements significantly bolster the model's capability to accurately detect intricate defect patterns. CNN: Convolutional Neural Networks, PANet: Path Aggregation Network, Transf: Transformer, E-PANet: Enhanced PANet.
• Figure 2: The proposed Fab-ASLKS model. Enhance the C2f module in the Backbone with the large kernel shift convolution (LKSC). Utilize the adaptive shape convolution (ASC) to expand the receptive field of the fourth C2f module in the Neck.
• Figure 3: The Framework of adaptive shape convolution (ASC). First, divide the $c$ channels into $g$ groups, then generate offsets ($\Delta m, \Delta n$) and weights ($w$) through $g$ group convolutions. Next, the offsets are added to the original convolution pixel positions, the convolution is performed, and the weights are multiplied at each position to obtain the ASC output.
• Figure 4: The Framework of large kernel shift convolution (LKSC). (a) Decompose the $51\times51$ convolution into three convolutions: $51\times5$, $5\times51$, and $5\times5$, using SLaK. (b) Decompose the $51\times5$ convolution into eleven $5\times5$ convolutions using shift convolution (SC), followed by fusion after shifting.
• Figure 5: Example images of 20 types of defects with ground truth bounding boxes