Real-Time Classification of Steel Strip Surface Defects Based on Deep CNNs

Abstract

Steel strip surface defects recognition is very important to steel strip production and quality control, in which correct classification of these surface defects is crucial. The surface defects of steel strips are classified according to various features, but it is hard for traditional methods to extract all these features and use them effectively. In this paper, we propose a method to deal with the problem of defect classification based on deep convolutional neural networks (CNNs). We adopt GoogLeNet, as our base model and add an identity mapping to it, which obtains improvement to some extent. At the same time, we establish a dataset of cold-rolled steel strip surface defects of six types and augment it in order to reduce over-fitting. Then we detect defects of six types with our network and reach an accuracy of 98.57%. Besides, our network achieves a speed of 125 FPS, which fully meets the real-time requirement of the actual steel strip production lines.