Research On PCB Surface Defect Detection Methond Based On Convolutional Netural

The development of Printed Circuit Board(PCB)plays a very important role in the development of the national economy.PCB surface defect detection is a hot and difficult problem in the PCB production process.Aiming at the problems of inaccurate positioning of PCB board defects and low detection accuracy of commonly used target detection algorithms,this paper proposes two algorithms to detect PCB board surface defects.The specific work is as follows:1.A method for detecting PCB surface defects based on Feature Pyramid Networks-Deformable Convolution Network(FPN-DCN)is proposed.By designing a feature pyramid module to fuse multi-scale features,the reference can introduce deformable convolution,and a variable convolution feature pyramid network structure is constructed.The advantage of this model is that the feature pyramid structure is added to fuse features to improve the model’s ability to detect defects;by adopting a variable convolution module,the model’s ability to detect morphological defects is improved.2.Aiming at the problems of inaccurate positioning of PCB board surface defects detected by the current common target detection algorithms,small defects are difficult to detect,this paper proposes the multi.scale feature fusion YOLO V3(Multiscale Feature Fusion,MFF-YOLO V3)PCB defect detection method.Inspired by the YOLO V3 model,multi-scale image features are extracted by designing a convolutional neural network,the generated multi-scale features are fused to generate single-scale image features,and then clustering methods are used to achieve accurate positioning of PCB board defects.The difference from YOLO V3 is that by improving the resolution and fusion of multi-scale image features,the model’s ability to detect small PCB defects is improved;in order to achieve precise positioning of PCB defects,k-means with Avg IOU as the gold standard is adopted.The algorithm realizes the redefinition of the candidate target area.At the same time,because MFF-YOLO V3 achieves a single output to achieve feature extraction,the number of convolutional layers is reduced,thereby reducing the amount of calculation for network training.Compared with general target detection algorithms,this method has higher accuracy for PCB board defect detection,and can improve the detection ability of PCB board surface defects in the industry.The two methods proposed in this article are trained and tested on the Deep PCB dataset.The result is that the detection accuracy of the FPN-DCN method reaches 89.93%,and the detection speed is about 20 frames/s.Compared with the Faster R-CNN algorithm,m AP is improved 5.4%;MFF-YOLO V3 has a detection accuracy of 87.9%,and its m AP is9.2% higher than that of YOLO V3.Experiments show that the two methods proposed in this paper are more effective for PCB surface defect detection,and basically meet the requirements of industrial detection.