基于YOLO-CH的PCB焊点缺陷检测算法OA

To address the problems of easy loss of small-target features,low detection accuracy,and excessive model parameters in the field of printed-circuit board(PCB)solder joint defect detection,a PCB solder joint detection algorithm named YOLO-CH based on an improved YOLOv8 algorithm was proposed.Based on the feature pyramid and path aggregation network of YOLOv8,a high-resolution P2 detection layer was added,reducing the downsampling stride to four.Combined with progressive feature fusion and bidirectional feature interaction strategies,key details such as texture and edges of small targets were effectively preserved.A channel and edge-wise context aggregation(CECA)two-stage attention module was designed,which adopts a cascaded structure of channel attention and coordinate attention to enhance the feature representation of small targets while keeping the model lightweight.A three-dimensional composite weighted loss function was designed to alleviate the imbalance in optimization between large and small targets.Experimental results on a self-built PCB solder joint defect dataset show that compared with YOLOv8 algorithm,the proposed algorithm improves the mAP50 by 4.5 percentage points,mAP50-95 by 4.6 percentage points,and recall by 5.3 percentage points,while reducing the floating-point operations(FLOPs)to 7.09×1010,effectively lowering model complexity and improving defect detection performance.