Research Based On Semi-supervised Manifold Learning Of Wafer Surface Defect Detection

With the globalization of semiconductor design and manufacturing processes,the requirements for defect detection in the manufacturing stage of integrated circuits will become higher and higher Traditional defect detection adopts manual use of electron microscope to collect sample pictures,and then submit to experts to assess whether there are defects.Therefore,it is necessary to design a model that can automatically detect defects indicated by experience instead of manual operation.Limited by the sample data and the type of detection in abnormal image detection,the existing data-driven supervised deep learning detection models such as YOLO and SSD often have greater limitations.This paper proposes a semi-supervised learning method.Its training method does not require negative label data,and only uses normal samples for training.The proposed method constructs a manifold learning model structure formed by encoder-decoder-encoder.First,the latent space vector is constructed using an autoencoder,which is used to obtain latent information of normal samples.Second,construct an adversarial generation network to strengthen the model's ability to reconstruct normal samples.Finally,the patch-GAN is cited to output the discriminator in the form of blocks to improve the model discriminating ability.In addition,a residual network model with a jump structure is added to the generated network to improve the model training effect.The method proposed in this paper only uses normal unlabeled data during training,and marks normal data and abnormal data in the detection phase to complete the model classification of the data.The model in this paper is verified on the benchmark data set Mnist,CIFAR10 and industrial data set WM-811 K wafer data and package X-ray data set.The experimental results show that the classification accuracy and detection effect are superior to the most advanced model method.The designed model meets the needs of industrial applications.