Lightweight Fabric Defect Detection Based on Improved YOLOv4

Abstract

Fabric quality is crucial for subsequent production, so automated detection of different types of defects on fabric surfaces is essential. To improve the applicability of the detection model in textile factories, a lightweight fabric defect detection method based on YOLOv4 is proposed, which firstly addresses the characteristics of large scale transformation of fabric defects and more small defects by combining the shallow features of the 11th layer in YOLOv4 with the deep features to generate a new scale feature layer; secondly, the SPP structure in YOLOv4 is improved to using pooling decomposition to improve the large pooling kernel and increase the model computation speed; finally, the depth-separable convolution is introduced to optimize the model size, making the model detection faster to meet the actual needs of textile mills. The improved network improves the defect detection accuracy, with an average detection accuracy of 79.31%, and its detection speed is also 5 times faster than that of the original YOLOv4, which is more suitable for the actual detection of textile mills than the original YOLOv4.