Machine Learning Based Defect Detection For Thermoelectric Coolers

With the increasing demand for efficiency and quality of semiconductor die production,manual inspection is no longer sufficient to meet the production needs,so how to effectively study a way to achieve automatic inspection of semiconductor die has become a current hot topic.Thermoelectric coolers(TEC),as one of the representatives,are prone to cracks,deformations and other defects in shape and size during the slicing and dicing process,and the existence of these surface defects can lead to a reduction in the life span and overall product performance.A machine learning-based defect classification and detection method for TEC images is proposed to ensure the quality of semiconductor refrigeration devices by strictly screening out qualified grains.In this paper,machine learning theory is applied to study and explore the issues of feature extraction,classification and detection of TEC images,and the main work is as follows:(1)A Gabor filter-based feature extraction method is proposed to achieve defect classification of TEC images.Firstly,twenty different Gabor feature maps are obtained by Gabor filtering from four different scales and five different directions of the image,and these twenty feature maps are used to represent the features of the original image to complete the feature extraction work of the image.However,due to the large dimensionality,in order to further reduce the computational difficulty and computational time for the subsequent classification,a principal component analysis method is introduced to reduce the dimensionality of the feature maps to obtain the final feature vectors,and finally the obtained feature vectors are classified by using different classifiers,and the experimental results show that this method can complete the image feature extraction well and achieve better classification accuracy.(2)A full convolutional neural network based on deep learning is proposed to achieve the detection of TEC defects.On the basis of Faster RCNN,Resnet50 is introduced in the feature extraction module to achieve feature extraction of the image in order to further better extract the features described in the image,and a feature pyramid(FPN)network is introduced to enhance the detection of small targets,which combines deep semantic information with the low-level detail information to achieve multi-scale detection of images.In the RPN module,the non-maximal suppression(NMS)is replaced with soft non-maximal suppression(Soft-NMS)in order to further reduce the missed detection rate of the model.The experimental results show that the proposed method in this paper can better detect the defect class and location of TEC.