关键词: YOLOv7, 织物疵点, 深度学习, 目标检测, 空洞卷积

Abstract: China’s current textile and clothing exports exceed one-third of the world’s proportion, ranking first in the world, so the development of textile enterprises is particularly important. But in the process of production, textiles are easily affected by environmental factors, equipment defects, human errors, poor technology, shortage of raw materials and other problems, resulting in skips, holes, water stains and other defects that affect product quality, and at present, as for the fabric defect detection link of most enterprises, manual visual inspection method with high cost and low efficiency is adopted, which causes worker fatigue and is easy to produce missed detection and false detection. Therefore, the research on intelligent fabric defect detection technology can effectively improve the production efficiency of textile enterprises, improve detection accuracy, and reduce production costs.
Compared with the traditional defect detection algorithm applied to fabrics, the detection method based on deep learning has better adaptability and learning. To solve the problems of low accuracy, high missed detection rate, slow training speed and difficult convergence of the model in traditional enterprise applications, based on the YOLOv7 algorithm with the latest Extended-ELAN architecture, the DCCSPC parallel stacking module was designed by using the cavity convolution with different parameter values, the SPPCSC spatial pyramid pooling layer was improved, and the local and overall characteristic information of fabric defects was deeply integrated. The top-down feature extraction network of the model was selected from the bottom of the feature extraction network, the feature matrix generated by the first ELAN module was spliced after 2-fold upsampling, a higher predicted feature map of 160×160×255 size was output, and the small defects of fabrics with a width and height of only more than four pixels were predicted. To solve the problem of slow convergence caused by increasing the number of model parameters, the CIoU loss function was replaced with the WIoU loss function, solving the problem of high missed detection rate of special samples (samples with aspect ratios inconsistent with most samples), and improving the convergence speed of the model. On this basis, 2,438 fabric defect image production datasets from the platform of Alibaba Tianchi were selected, and through ablation tests and comparative experiments with other detection algorithms, it was shown that the improved model could effectively detect hair defects with a lower proportion of pixels, predict defects such as dead wrinkles in thin strips, improve the recognition rate of water stains and other color and background defects, and increase the average accuracy value by 3.4%, which can meet the conditions of industrial-grade deployment and real-time detection, and be more effectively applied to intelligent defect detection in industrial production.    
Through the specific characteristics of fabric defect types, the YOLOv7 algorithm based on deep learning is used to improve the detection accuracy of specific defects and small defects, which provides an effective method for the detection of defects in intelligent production of textile enterprises, improves the detection quality, and provides an effective idea for further research on the defect detection algorithm.

Key words: YOLOv7, fabric defects, deep learning, object detection, dilated convolution