Metal Surface Defect Identification Method Based on Deep Learning

Abstract

Aiming at the problems of various defect types, sizes and shapes in the process of identifying metal product surface defects, a deep learning model based on multi-scale residual convolutional network is proposed. The network uses ResNet50 as the feature encoder to extract feature maps with different resolutions to capture multi-scale feature information, thereby improving its ability to identify defects of different sizes; it also uses Multi-Layer Perceptron (MLP) for multi-scale Adaptive fusion of features enables information interaction and feature refinement between features such as image texture and boundaries obtained by shallow convolution and complex semantic feature information extracted by deep convolution to improve network model recognition performance. Experimental results show that the algorithm proposed in the article has an accuracy of 98.06% on the NEU-DET data set, and has higher recognition accuracy than other models.