Classification of Steel Strip Surface Defects Based on Optimized ResNet18

Abstract

Steel strip is one of the main products of traditional steel manufacturing enterprises. It is of great significance to accurately identify the types of defects on the surface of the steel strip. This paper innovatively proposes a pre-training method of network weights based on ResNet18. The network is optimized by dynamically adjusting the learning rate. This method can classify steel strip images with high accuracy of 98.585%, avoid overfitting and enhance the stability of training process.