Research On Key Problems Of Electronic Components Defect Detection Based On Deep Learning

In the production of electronic components,the defect detection process is critical to the quality of electronic components.Defective products need to be detected accurately to ensure the product yield rate of electronic components.Currently,defect detection methods based on machine vision are widely used in practical production.However,traditional machine vision methods for defect detection of electronic component images depend on a large number of features that are extracted manually.For defective images with noise and low difference from non-defective images,traditional machine vision methods are usually difficult to meet the requirement of accuracy in practical applications.Compared to traditional machine vision methods,deep learning methods have outstanding performances because of the ability to extract deeper and more complex features from images.Thus,this paper proposes image defect detection methods based on deep learning.In addition,this paper conducts further research and discussion on the following actual situations in defect detection of electronic components: Some defects in images of electronic components to be detected are relatively minor;there are many kinds of defects in electronic components,but it is difficult to obtain all kinds of defective samples in actual production process;the actual number of defective samples obtained is much less than that of non-defect ones,which causes the problem of imbalanced samples;a large number of parameters make it difficult to balance efficiency and accuracy in deep learning.Focusing on the above problems,the main research results of this paper are as follows:(1)Due to the insufficient diversity of defective samples in electronic components dataset,the distance between classes is not much larger than the intra class distance(i.e.distance is inseparable),which makes it difficult to determine the boundary of defective class.It accords with characteristics of one class problem.Thus,a one-class convolutional neural network(CNN)model is proposed in this paper for the defect detection of electronic components.By setting the classification threshold,the model implements the clustering of non-defective(positive)samples,which can improve the classification accuracy when the diversity of negative samples is insufficient.In order to verify the performance of the classifier,a general two-class convolutional neural network and CNN-SVM is designed as comparison.Experiment results show that the proposed one-class CNN performs well on this datasets.(2)Aiming at the imbalance of positive and negative samples in images,a salient region oversampling method for defective images is proposed to sample datasets according to the characteristic of defective images.The proposed method increases the number of defect samples and solves the problem of imbalance between positive and negative samples.In order to verify the validity of the over-sampled images,the one-class CNN and two-class CNN are designed and be tested on the dataset.Experimental results are compared with those of under-sampling and repeated over-sampling,which show that the oversampling method can improve the detection accuracy.Also,the results indicate that the balanced samples obtained by the over-sampling method can improve the defect detection accuracy to a certain extent.The two-class CNN is shown to have higher classification accuracy and efficiency than one-class CNN on large image datasets.(3)In order to improve the accuracy and efficiency of defect detection,a Two-Step Convolution Neural Network(TSCNN)model is proposed for large image sets with imbalanced samples.First,it classifies the original large dataset by two-class CNN model.Image with probability less than 0.25 is regarded as negative class and image with probability greater than 0.75 is classified as positive class.Images with probabilities between 0.25 and 0.75 need to be reclassified by one-class CNN because of the lack of confidence.Compared with general convolutional neural networks,TSCNN improves the accuracy by 1.37% while ensuring efficiency.(4)In order to further improve the accuracy and efficiency of defect detection,this paper proposes the improved Binary Neural Network(BNN)based on adaptive pooling for defect detection,which can compensate for the loss of accuracy caused by quantization,and is able to speed up computing operation.Compared with the general BNN,the accuracy of the binary network based on adaptive pooling is improved by 0.87% without increasing operation time.