An Improved and Efficient PSMD-Yolo Algorithm for Detecting Surface Defects in Fabrics

Addressing the issue of poor accuracy in traditional detection methods for fabrics with significantly varying textures and small defect targets, this paper proposes an improved detection algorithm, PSMD-Yolo, based on YOLOv8. Firstly, a lightweight convolution module, PP-HGNet, enhancing the network's ability to process complex image data. This is combined with the S2-MLP attention mechanism to fuse the segmented feature maps, thereby improving image recognition accuracy. Secondly, MAFPN is added to the Neck layer to increase the efficiency of feature fusion . Finally, the MPD-IOU loss function is introduced, which considers the distance between center points to achieve more accurate regression results. Experiments were conducted on a fabric defect dataset collected from a textile factory in Zhejiang. Compared to the original YOLOv8 model, the P-value increased by 1.4%, the R-value by 2.7%, mAP50 by 1.4%, and mAP50:95 by 2.8%. The detection P-value and mAP50 accuracy reached 96.5% and 98.1%, respectively. The average precision improved by 6.5, 4.8, and 1.4 percentage points relative to SSD, Faster-RCNN, and Centernet, respectively.To further verify its generalization, the algorithm was evaluated on the Alibaba Cloud Tianchi fabric defect dataset, resulting in an average precision increase of 3 percentage points.