Fabric defect detection network incorporating complementary-strategy downsampling and global information fusion

Abstract In the defect detection of textiles, the size differences of the defects pose a significant challenge for accurate identification, especially under the constraints of lightweight models. This paper aims to propose a textile defect detection network, using a complementary strategy of downsampling module and directional global information fusion module to address the problems of information loss of small sizes during downsampling and the modeling difficulty of large-sized defects. Specifically, a complementary strategy downsampling module is designed, which integrates the continuity retention based on interpolation, the structure retention based on slicing, and the robustness based on pooling into a unified downsampling framework to replace the convolutional downsampling in deep learning networks, in order to retain the key small-scale and boundary information. A directional global information fusion module is constructed, which models the long-range dependencies along the warp and weft directions of the fabric to enhance the model’s ability to model large-sized defect images with strong directional features. Moreover, in order to achieve a balance between enhanced feature representation and computational efficiency, a lightweight detection head based on shared convolution is adopted to significantly reduce parameter redundancy while maintaining detection capability. Experimental results demonstrate that the proposed network provides a good balance between accuracy and efficiency, achieving 82.4% detection accuracy with 2.453 M parameters and 7.1 GFLOPs, along with an inference speed of 182.4 FPS. Compared with the most advanced method, the detection accuracy is improved by 3.39%, making it more suitable for practical industrial applications.