MWBD-YOLO: a multi-wedge belt defect detection method with hybrid architecture of CNN and transformer

Abstract Appearance defect detection is a critical task for ensuring the quality of Multi-wedge belt products. However, traditional convolutional neural network (CNN)-based defect detection methods face challenges such as weak global feature extraction and limited multi-scale feature fusion capabilities, which affect detection efficiency. To address these issues, we propose a high-precision defect detection method for multi-wedge belts, MWBD-YOLO (Multi-Wedge Belt Detection), which combines CNN and Transformer. Firstly, we introduce a hybrid CNN-Transformer network with adaptive feature extraction, incorporating the lightweight Swift-Former module into the CNN structure. This fusion enhances the model’s ability to capture long-range semantic information by leveraging the strengths of both CNN and Transformer. Secondly, a multi-scale feature fusion network is incorporated into the neck of the model, through the design of a multi-scale path aggregation network and the introduction of a dynamic upsampling module, Dysample. This allows the model to more effectively capture features at different scales, addressing the challenges posed by feature scale differences. Thirdly, we introduce a new loss function, Wise-IoU, to reduce the impact of low-quality instances on the boundary box loss function in object detection. This dynamic non-monotonic focusing mechanism, based on WIoU loss, focuses on anchor boxes of ordinary quality, improving the model’s defect detection performance. Finally, we determine the optimal lighting method and design an image acquisition system for capturing defect images on the production site, creating a defect dataset. This dataset, named MWB-DET, contains a diverse set of images covering six major defect categories: concavities, impurities, scratches on the tooth surface, and bumps, cords, and folds on the back of the belt. Experimental results on the dataset demonstrate that the proposed algorithm achieves a mAP50 of 94.5%, precision of 92.9%, and recall of 91.0%. Compared to the baseline model, the proposed method improves mAP50 by 4.5%, precision by 3.8%, and recall by 6.6%. The algorithm demonstrates superior accuracy and speed compared to other advanced methods, achieving excellent detection results. This method meets the industrial scene requirements for accuracy and real-time performance and has been successfully deployed in real-world scenarios. It provides technical support for defect detection and production quality control.