【摘要】 针对目前PCB缺陷检测方法中存在缺陷较小不易识别、缺陷形状多样化导致识别率下降等问题,提出基于多尺度特征融合和可变形卷积的PCB缺陷检测算法(DCR-FRNet)。在Faster R-CNN算法的基础上进行优化改进,能够更好地适应同一缺陷不同尺度的缺陷目标。采用的多尺度融合的金字塔模型有效地提高模型的特征识别能力;引入的可变形卷积替代常规的卷积,通过卷积学习偏移量提高模型的特征提取能力。实验结果表明,在采集的缺陷数据集上,所提DCR-FRNet算法相对于基准网络能够更有效识别缺陷特征,检测精度达到了96.60%,F1分数提高了16.30%。更多还原

【Abstract】 Aiming at the problems of current PCB defect detection methods that are difficult to identify small defects,and the problem that its recognition rate is reduced due to the diversification of defect shapes,a PCB defect detection algorithm(DCRFRNct) based on multi-scale feature fusion and deformable convolution was proposed.The optimization and improvement based on the Faster R-CNN algorithm better adapted to the same defect and different scale defect targets.Among them,the multi-scale fusion pyramid model effectively improved the feature recognition ability of the model.The introduced deformable convolution was used to replace the conventional convolution,and the feature extraction ability of the model was improved through the convolution learning offset.Experimental results show that,on the collected defect data set,the proposed DCR-FRNct algorithm can identify defect features more effectively than the benchmark network.The detection accuracy reaches 96.60 %, and the F1 score is increased by 16.30%.更多还原