ABSTRACT Industrial products play an indispensable role in intelligent manufacturing and the development of high‐end equipment, but they create challenges in the production process such as similarities between different types of defects, weak semantic information of complex small targets, and multi‐scale changes in defect targets. To address the above problems, we propose a novel U‐shaped symmetrical interactive fusion network (USIF‐Net). First, a local‐global feature extraction network (LGFE‐Net) is proposed to solve the defect diversity and similarity problems, maintaining local detail feature extraction while strengthening global context modeling, enabling it to accurately capture defect‐related features. Second, a symmetrical U‐shaped progressive interactive convergence network (PIC‐Net) is designed, which adopts a bidirectional feature aggregation mechanism to achieve deep feature interaction between adjacent levels and cross‐levels. A path aggregation module (PAM) is proposed in the neck network to interact with features from different levels, ensuring that fine information is gradually recovered during feature transmission and improving multi‐scale defect detection. Finally, an adaptive feature fusion module (AFFM) is proposed to filter cross‐level conflict information and effectively solve the detection performance degradation problem caused by semantic gap in feature pyramid networks. The experimental results show that mAP@.5 of 94.7%, 98.8% and 97.1% were obtained on the NEU‐DET, PCB and AL surface datasets, respectively. Compared with the baseline model, they improved by 4.4%, 1.6% and 1.7%, respectively, and reached 52FPS, achieving SOTA performance in comparison with mainstream models, providing reliable technical guarantees for intelligent manufacturing quality control.
Zhang et al. (Thu,) studied this question.