Research On Inductive Components Defect Recognition Based On Deep Learning

With the gradual miniaturization and high density of electronic assembly,electronic packaging testing technology has become more and more important.In order to improve production efficiency,many enterprises have introduced assembly line welding machines such as chip mounter,but the workpieces produced by automatic welding machines have a certain defect rate,such as less tin,more tin,bridging and tin tip.At present,computer vision is widely used in the electronic packaging and testing industry.It mainly judges defects and faults through computer processing,analysis and comparison.Its biggest advantage is lower cost and higher production efficiency.The unified testing standard can eliminate the interference of human factors,ensure the reliability,repeatability and accuracy of testing results,and effectively improve the product quality and automation level of the enterprise.Traditional computer vision technology generally uses image filtering,binarization and morphology to reduce noise,and uses manual features such as gray,contour and color for image matching.These concise technologies can be competent for some simple scenes,but when facing the situation that the background and environment are more complex and the target characteristics are more diverse,the traditional computer vision technology is difficult to achieve enterprise level application accuracy.Moreover,when the detection target changes,the corresponding manual features also need to be redesigned and evaluated.As we all know,the quality of features has a vital impact on the generalization of detection,and it is not easy to design good manual features.In recent ten years,deep learning has gradually dominated the field of computer vision.It mainly produces good features through machine learning technology itself,which is very suitable for complex defect targets such as high-frequency inductive solder joints.The existing computer vision systems often use vertical single view detection in electronic packaging testing.For some products whose three-dimensional defect shape needs to be judged by combining multiple faces,it can not be well considered.The main bottleneck lies in the extraction,description and matching of target multi-dimensional features and the performance of related algorithms.Therefore,based on the multi view detection structure,this thesis studies the related technologies of multi-dimensional feature recognition and defect location of solder joints.The main research contents of this thesis can be summarized as follows:(1)According to the functional requirements of the vision system,the multi view vision detection system framework and core module are designed.In view of the conical shape of the target solder joint,the framework uses the left,right and top three perspectives to obtain the surface information of the target image,and designs the multi view solder joint data set according to the framework.(2)Aiming at the problem that the overall features of the target are ignored in the process of feature extraction of the prototype network,the crossed network module is designed,and a multi view joint adaptive learning framework based on cross network is proposed.The results show that the framework shows better accuracy and reliability than single view.(3)Using the multi-scale target recognition algorithm based on cross network proposed in this thesis,the industrial field application of high-frequency inductive solder joint defect recognition is realized.The test shows that the production efficiency of this kind of products has been improved and online production has been realized.