基于改进的RT-DETR小目标PCB缺陷检测OA

An improved RT-DETR algorithm is proposed to tackle common challenges in printed circuit board(PCB)defect detection,including missed detections,false detections,and the low detection rate of tiny defects.Firstly,a multi-scale feature edge information fusion(MSFEIF)module is designed in the backbone network to enhance the network 's ability to extract multi-scale edge features and achieve global propagation of edge information.Additionally,an efficient adaptive attention(EAA)mechanism is introduced to the original AIFI module,and an AIFI-EAA scale interaction module is constructed.This modification reduces computational complexity,speeds up inference,and improves the detection accuracy of the model.Furthermore,combining the advantages of RepC3 and RmtBlock structures,a new RmtBlockC3 module is designed to effectively enhance the spatial modeling and global perception capabilities of the network.Finally,by integrating the Inner-IoU concept with the MPDIoU loss function,the matching accuracy of bounding box regression and keypoint localization accuracy is improved without increasing model parameters,so as to further enhance the detection performance.Experiments on the PCB defect dataset demonstrate that the improved model achieves superior performance compared to the baseline RT-DETR.Specifically,its precision increases by 4.3%,its recall rate by 3.3%,and its mean average precision(mAP@0.5)by 3.5%.These results indicate that the proposed algorithm holds high value for the practical application of small object defect detection in PCB inspection.