Research On Solder Joint Detection Algorithm Based On Machine Vision

With the development of electronic manufacturing industry,solder joint detection technology has been widely used.The quality and reliability of solder joints affect the quality and efficiency of the finished products.Therefore,how to improve the quality of solder joints has become a hot issue in the electronics industry.In recent years,machine vision as an emerging detection technology has been used widely,greatly improve the detection accuracy and efficiency.But in the process of the solder joint detection based on machine vision still exist two problems:First is the most classifier using simple Euclidean distance measurement features as the input vector of the differences between.However,Euclidean distance measures cannot use any statistical rules for the data of known training sets.Second is the poor effect of classification algorithm commonly used in machine vision of nonlinear sample,resulting in detection rate and missing problem.Therefore,this thesis,based on the machine vision circuit board solder defect detection as the research background,feature extraction and classification algorithm as the theoretical basis,Classify the tin solider as normal solder,more solder and less solder.The main work of this thesis is as follows:(1)Analysis and research metric learning,classification algorithm and its improvement.By comparing the BP neural network,the K nearest neighbor method and support vector machine(SVM)three algorithms respective superiority and practical application scope,K nearest neighbor algorithm is selected as the basis of classification algorithm.According to the Euclidean distance which is used in the K nearest neighbor method,the statistical rules cannot be used.The metric learning method is introduced to study the characteristics of statistical learning,to find out the law of the data,and to reflect the importance of different types of features.And for the problem of nonlinear classification effect,by transforming the sample from low dimensional space to high dimension space,the improvement of K neighbor method is realized.Finally,the improved K nearest neighbor method based on the large boundary is formed.(2)Positioning and feature extraction of solder joints in circuit board image.Use industrial camera to collect the solder joint image of the circuit board and adaptive median filtering processing images.Then based on the shape of the template matching method to locate the image of the solder joints,and according to the characteristics of the solder joint distribution histogram extracted features and normalized processing.Finally according to the distribution of the solder joint characteristics,the characteristics of the solder joiint types can be clearly distinguished as an effective feature.(3)Construct defect detection classifier model.The representative training samples are selected to extract the effective sample features and the normalization process is used as the input of the detection model.We use the improved K nearest neighbor method based on large boundary to study the characteristics of the training samples to form the solder joint defect detection classifier model,and the model is tested by the test samples.(4)Solder joint defect detection and contrast experiment.The thesis compares the algorithm of this paper with BP neural network,the K nearest neighbor method and support vector machine(SVM)algorithm in the Solder joint detection experiment.By comparing the accuracy rate of solder joint inspection,false detection rate,the missing rate,and detection time,finally it is concluded that the algorithm can effectively solve the surrounding noise and nonlinear problem of poor effect of sample classification.At the same time,the algorithm makes full use of the rules of the characteristics,so as to improve the detection accuracy and reduce the testing time.