Research Progress in Deep Learning for Ceramics Surface Defect Detection

Man Zhou,Tianzhao Wu,Zhongyi Xia,Bin He,Ling Bing Kong,Haibin Su

DOI: https://doi.org/10.1016/j.measurement.2024.115956

IF: 5.6

2025-01-01

Measurement

Abstract:Because the surface defects of ceramics are widely varied, traditional machine vision algorithms are not sufficiently effective in achieving complete modeling of defective features, while the reusability is not large and the working conditions should be differentiated, thus leading to huge labor costs. Deep learning in feature extraction and localization has been intensively used to study surface defects of ceramics. With a suitable dataset and appropriate network models and algorithms, combined with subsequent continuous training of the neural network models, the automatic detection and classification of ceramics surface defects could be realized. Current commonly used deep learning surface defect detection algorithms include Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Multilayer Perceptron (MLP), etc. Because of its specific characteristics, CNN can be used to directly detect and classify two-dimensional surface defects. In this case, more abstract features from the original image can be extracted by using a simple nonlinear model, with very low human labor involvement in the whole process. In this paper, CNN was selected as the focus, from four aspects to study the application of CNN-based and CNN-derived neural network models in ceramic surface defect detection. The first step is to select a suitable neural network model for surface defect detection, such as YOLO, MASK-RCNN and other neural networks, which are widely used for image classification, target recognition, speech recognition and so on. In second step, the samples needed to train the neural network model are obtained and then preprocessed. In third step, appropriate feature extraction algorithm and sample training algorithm are selected, model evaluation criteria are formulated and the effects of neural network models on the performance of surface defect detection are compared. In the last step, the prospects of deep learning for surface defect detection and the direction of improvement are summarized and discussed.