DC-YOLO: A dual channel YOLO for small object defect detection of circuit boards

In response to the limitations of existing circuit board defect detection methods, including low detection efficiency, severe leakage of small-sized defects, and incomplete feature capture, this paper proposes an enhanced defect detection network called DC-YOLO based on YOLOv7. Firstly, a spatial pooling pyramid, DC-SPPFCSPC, is constructed in this paper. It combines multi-dimensional salient features with global spatial features by extracting and locating a dual-channel structure. The enhancement focuses on establishing the measured connection between defective points and the global features while preserving the spatial attributes of deep image features. Secondly, the model improves the FPN by fusing higher dimensional features. This enhancement enriches the shallow feature information without altering the original number of detection heads. Consequently, this method retains more relevant information while avoiding increasing the burden on the detection head. Lastly, CBAM is integrated into the FPN input. This integration enhances the algorithm’s ability to concentrate on the defective area, providing more detailed feature information. The experimental results indicate that the mAP of the enhanced algorithm reached $98.39\%$ and $98.56\%$ respectively on the public dataset and the dataset with defective homemade circuit boards. There was a significant improvement in small object detection.