76‐2: Defect Classification Algorithms for Display Manufacturing Based on the Convolutional Neural Network Mixture‐of‐Experts Model

In the field of industrial display manufacturing, a large number of proprietary defect identification models for detailed scenarios have generated due to the differences in production lines, environments and defect types. The sparse network structure was designed to classify defects by studying the method of mixing these pre‐expert models: the model fused the features extracted by the pre‐expert models through the gating unit, and the Mixture‐of‐Experts(MoE) layers were used to extract and transform the semantic features, andfinally the prediction results were obtained through the sparse decoding structure. In order to prevent overfitting caused by a small amount of data, this paper improved the convergence performance of the model and reduced the computational cost by reusing parts of the pre‐expert models to improve the generalization of feature extraction. This paper provided more possibilities for the defect classification tasks of the MoE model based on Convolutional Neural Network(CNN).