CM-YOLO: A Multimodal PCB Defect Detection Method Based on Cross-Modal Feature Fusion
Haowen Lan, Jiaxiang Luo, Hualiang Zhang, Yan Xu
Abstract
By integrating information from RGB images and depth images, the feature perception capability of a defect detection algorithm can be enhanced, making it more robust and reliable in detecting subtle defects on printed circuit boards. On this basis, inspired by the concept of differential amplification, we propose a novel and general weighted feature fusion method within the YOLO11 dual-stream detection network framework, which we name CM-YOLO. Based on the differential amplification approach, we introduce a Differential Amplification Weighted Fusion (DAWF) module, which separates multimodal features into common-mode and differential-mode features to preserve and enhance modality-specific characteristics. Then, the SE-Weighted Fusion module is used to fuse the common-mode and differential-mode features.In addition, we introduce a Cross-Attention Spatial and Channel (CASC) module into the detection network to enhance feature extraction capability. Extensive experiments show that the proposed CM-YOLO method achieves a mean Average Precision (mAP) of 0.969, demonstrating the accuracy and effectiveness of CM-YOLO.