Study On The Key Technique Of PCB Defect Detection Based On High-Resolution Images

This paper studies on the algorithms of PCB defect detection based on high resolution images in PCB defect detection system.Such algorithms compare the tested image,which is sampled by a linear array CCD,with a defect-free template image to detect PCB defects online.The detected objects are six common PCB defects including open,short,mousebite,spur,copper and pin-hole.The overall task can be separated into two individuals: image registration and defect detection.The performance of image registration between the tested and template image pair has direct influence on the efficiency and accuracy of defect localization and classification.In this paper,we firstly register tested image onto the template through projection transform according to the coordinates of mark points on the both images.As for the distortion caused by frame loss or the irregular scan speed of the CCD,a registration algorithm based on interpolation of K-nearest neighbour and radial-based function is proposed,which calibrates the tested image by estimating the offset according to the displacement of K-nearest feature points.As for the distortion due to the deformation of the PCB itself,thin plate transform is adopted to further calibrate the tested image.According to the experiment results,most of the misalignment area between the registered tested and template images is smaller than 3 pixels,which satisfies the requirement of the following localization and classification algorithm.The PCB defect detection task contains two sub-tasks: localization and classification.An algorithm based on image processing and connected component analysis is proposed to locate and mark each suspected area of PCB defect in the tested image.As for classification sub-task,firstly,a soft-margin kernel SVM is deployed as a baseline to classify PCB defects.As for the problem that traditional feature extraction algorithms,e.g.HOG,SIFT,can hardly help the classifier to achieve high performance,secondly,three new feature extraction methods for PCB defect classification are proposed: the consistency of connected component,the consistency of border,and the mean and variation of Hausdorff distance of the edge points.Relying on such three features,a minimum risk sequential classifier is proposed to classify those defects.In the meanwhile,the method for estimating the parameters in the classifier is given in this paper.To deal with the heavily dependent on manually feature selection,the insufficient generalization ability and the ignorance of tiny defects,thirdly,a model of deep neural network for PCB defect detection is designed to simultaneously locate and classify PCB defects according to simply registered tested and template image pair.In this model,a novel group pyramid pooling module merges convolutional features at various resolutions,which are then used to detect PCB defect in correspondingly sizes.In order fairly and efficiently evaluate those PCB defect detection algorithms and models,we provide a new dataset,namely DeepPCB,which contains 1,500 pairs of aligned tested and template images with size of 640×640 pixels.The experiment results show that the proposed localization algorithm can locate the PCB defects with the size greater than 3 pixels.The proposed minimum-risk based sequential classifier can recognize the six types of PCB defect precisely,which relies on the proposed three feature extraction algorithms.Moreover,the PCB defect detection model based on group pyramid pooling structure can achieve 98.6% mAP on the test set of DeepPCB with high efficiency of 70 FPS,which surpass the popular object detection models,e.g.,YOLO,SSD and Faster R-CNN,in PCB defect detection task.It also shows the power to detect tiny defects which are ignored by the methods based on traditional image processing algorithms.