Space-depth mutual compensation for fine-grained fabric defect detection model

Abstract

In recent years, using the deep learning approach in the textile industry for defect detection has emerged as a prominent research. However, detecting fabric defects remains challenging due to the small size and small number of fabric defect features. Traditional down-sampling operations that result in loss of feature information, interpolation up-sampling operations that add a lot of background redundant information, and interference with fabric images from external sources such as lighting or electromagnetic devices are significant barriers to achieving accurate defect detection using existing methods. In this work, we introduced a lightweight fabric defect detection method with enhanced resistance to interference. Firstly, we use YOLOv7-tiny as the basic model and integrate the Spatial Pyramid Dilated Convolution (SPD) and Efficient Channel Attention (ECA) modules to enhance the original MP-1 and Effective Long-Range Aggregation Network (ELAN) modules to retain fine-grained information, solve the problem of down-sampled feature loss and improve feature importance allocation. Secondly, a distinctive up-sampling Module (DTS) was proposed to replace the traditional interpolation up-sampling. The module expands the feature map size without adding extraneous information, thus ensuring more efficient integration of features of different sizes. Finally, a novel noise filtering technique called the Color Space Iterative (CSI) method was proposed to filter noise interference quickly and conveniently. Experiments on the open-source DAGM and TILDA defect datasets, as well as supplementary tests on CIFAR10 datasets for the CSI method, have yielded promising results. With a mere 3.4M parameters, the proposed lightweight model underscores the method’s superiority over the baseline in balancing model parameters, detection speed, and accuracy.