SDLS-YOLO: An ultra-lightweight real-time fabric defect detection algorithm based on multi-module coupling

Current techniques for detecting fabric defects frequently encounter challenges such as com-plex structural requirements, high computational demands, and inherent speed-accuracy tradeoffs. To address these limitations, this study introduces SDLS-YOLO, an ultra-lightweight real-time detection algorithm for fabric defects based on a multi-module coupling architecture. Firstly, the C2f-Star structure is engineered to extract fine-grained characteristics from images. Utilizing a multi-branch feature extraction technique, it effectively captures delicate textures and anomalies on the fabric surface. Secondly, an ultra-lightweight dynamic upsampling (DySample) module replaces the original upsampling component, enhancing the model’s capacity to detect fabric defects. Thirdly, the self-developed Lightweight Multi-Scale Fusion Module (LMSF) employs depthwise separable convolutions and feature reuse principles to facilitate information exchange between features of different scales in the deep semantic space, reducing computational complexity while significantly improving detection accuracy. Finally, a composite loss function, Shape Distance Adaptive Focaler IoU (SDAF-IoU), is developed to comprehensively account for the geometric characteristics and shape variations of fabric defects, enhancing the model’s sensitivity in detecting fabric defects and effectively diminishing the rates of misdetection and omission. Experimental findings indicated that the SDLS-YOLO model attains an F1 score of 93.5% and mAP50 of 96.7%, with a model size of only 4.8 MB and a frame rate of 185 FPS, satisfying the real-time demands of textile industry applications. In addition, the model’s generalization potential is confirmed using public datasets. Subsequent research focuses on refining the model using real manufacturing feedback data.