The detection of fabric defects is an important aspect of textile quality management. This paper proposes an algorithm based on YOLOv3 and deformable convolutional network to solve the problems of low accuracy, high rate of missed and false detection, and high labor cost existing in traditional manual detection methods. First, data enhancement is adopted to address the problem of imbalanced categories of defective samples in the dataset. Then, the Resnet101 model is used to extract the features, and the original convolution operation is replaced by deformable convolution to balance the accuracy and speed of the model. Finally, the focal loss function is introduced to solve the insensitivity problem of the model to difficult defect samples. Experimental results show that this method can quickly and accurately detect defects on the surface of fabrics. In real-time detection, the average accuracy of this method is 8.3% higher than that of YOLOv3. The average accuracy of multiple categories of detection reaches more than 90%, and the detection effect of small defect targets, such as three silk, is also relatively better.

Defects on fabric surfaces are difficult to identify owing to unsuitable computing devices, highly complex algorithms, small size, and high degree of integration with the fabric. To this end, this study proposes a lightweight fabric defect-detection network, YOLO-SCD, based on attention mechanism. The introduction of depth-wise separable convolution and the attention mechanism enhanced the capacity of the neck network to extract the defective features and increased the detection speed of the overall network. The extensive experimental results revealed that YOLO-SCD achieved an average accuracy of 82.92%, effective improvement of 8.49% in mAP, and an improvement of 37 fps compared to the original YOLOv4 on a standard fabric defect dataset. By leveraging its swift detection speed and high efficiency, YOLO-SCD excels in both the general fabric defect category and the difficult-to-detect fabric. Overall, it exhibited strong performance in detecting both minor flaws and flaws with high fabric integration. Furthermore, the proposed model was extended to steel datasets with similar characteristics.

No abstract