Research On Surface Defects Detection Of Micro Parts Based On Convolution Neural Network

With the development of industry,especially the progress of nanotechnology and integration technology,the size of parts becomes smaller and smaller so that the requirement for the accuracy and speed of the quality detection for micro parts becomes higher and higher.Nowadays,manually detection with microscopic tools is also the most commonly detection method,which may cause some deviations.In addition,low efficiency,high work intensity and low accuracy are also the drawbacks of traditional detection methods.Therefore,a more efficient and accurate detection method is needed to achieve the automatic detection of micro parts.Computer microscopic vision be introduced into the system based on the analysis of the current situation and requirements of the detection system,and the target detection algorithm based on convolutional neural network be used to detect the defects of micro parts.According to the actual conditions,on the basis of rationality analysis,the circumferential solder pad of PCB circuit board is selected as the testing object.The defects detection system based on convolutional neural network is designed,and surface images of micro solder pad are collected by microscopic vision to build dataset which the objective detection network is trained on.Finally,the trained network is used to make defects detection.The hardware system and software system be designed or chosen which includes the hardware type,module composition and working flow based on the actual requirement.Convolutional neural network(CNN)is the basic network of defects detection,which is used to extract the deep features of the defect images,and the features are used to distinguish the defect types.The network structure,data propagation method,and weights updating algorithm,etc.are deeply studied so that the improved scheme can be proposed in this part.The scheme be verified and tested by training on a new designed network,at same time,the rationality of applying CNN into defects classification system can be verified.Based on the research of CNN,Faster R-CNN is chosen as the core algorithm of surface defects detection for micro parts.The principle and implementation method of Faster R-CNN are deeply analyzed.Based on the experimental results,an improved scheme of network structure is proposed,and an experiment is designed to verify the target detection performance of the improved network.Finally,The defects detection network is trained by using the collected part defects dataset with the improved algorithm and structure based on Caffe.Applying the obtained detection network model to the actual defect detection task,the accuracy and speed of the detection method and the rationality of the method can be tested.