Research on Steel Surface Defect Algorithm Based on Deep Residual Network

This paper aims at the problems of hot-rolled-steel quality in industrial manufacturing and the difficulty of manual identification, low efficiency, and health hazards. We propose an end-to-end recognition method based on deep residual network to realize the automatic classification of hot-rolled steel surface defects. This method can effectively improve the production efficiency of hot-rolled steel. For the problem of insufficient negative sample data sets in the industrial field, we use varieties of data enhancement strategies to expand the original data, which solves the phenomenon of over-fitting due to insufficient samples during the model training process. The defect features are extracted through the CNN layer. Moreover, the residual structure is introduced to solve the problem of gradient disappearance and degradation when the network layer is deepened. The experimental results indicate that the accuracy of the ResNet-50 network model on the hot-rolled steel defect test sets can reach 93.34%, which is higher than the accuracy of the traditional network model. It also demonstrates this method has high reliability in the identification of defects that often occur in hot-rolled steel processing, including Rolled-in Scale, Crazing, Inclusion, Patches, Pitted Surface, and Scratches. The method proposed in this paper can meet the demand for industrial identification in the production process.