Research On Thermoelectric Coolers Wafer Defect Inspection Based On Deep Learning

Thermoelectric Coolers(TEC)are widely used for electronic cooling in a variety of applications ranging from consumer products to aerospace equipment,and are composed of many semiconductor wafers arranged.These wafers surface may have various defects after going through some complex processes,so it is very necessary to study a solution to this problem.In recent years,the advantages of deep learning technology in the field of computer vision are obvious.Therefore,this thesis considers using the deep learning method to solve the problem of TEC wafer defect inspection,and proposes two different detection methods according to the actual industrial needs.The main contents are as follows:(1)For online detection tasks with high real-time requirements in industry,the thesis proposes an improved YOLOv3 algorithm.Structurally,the algorithm adopts the residual network Res Net-D as the backbone network to reduce the redundant network parameters of the model.Optimize the regression of bounding boxes in the model with Complete Io U loss function.In the training process,the optimal clustering algorithm K-means++ is proposed to generate the anchor box of the predicted position,which can better fit the position of the TEC wafer defect.In addition,a variety of data augmentation methods are used to expand the dataset for the problem of insufficient pictures of the collected TEC wafer samples.The experimental results show that the detection accuracy of the improved model is increased by about 5%,and the defect detection speed is significantly improved.The FPS value reaches 83,and the memory occupied by the model is greatly reduced by more than 30%.(2)Offline detection pays more attention to the accuracy and stability of the detection method.This thesis proposes an anchor-free algorithm FCOS-EAM that integrates the attention mechanism.The model uses the efficient backbone VOVNet to extract image features and fuses the combined attention module CBAM of channel and spatial on its basis,which further strengthens the feature representation ability of the network.Finally,for the post-processing operation of filtering redundant prediction boxes,the soft-NMS algorithm is used for optimization.Experiments show that FCOS-EAM has a more efficient and accurate detection effect,and the detection accuracy reaches more than 90%.