Slighter Faster R-CNN for real-time detection of steel strip surface defects

Effective surface defect detection methods are of great significance for the production of high quality steel strip. Aiming at real-time detection of steel strip surface defect, this paper constructed a slighter Faster R-CNN. Firstly, convolutional layers for feature extraction in Faster R-CNN were replaced by depthwise separable convolutions so that the speed of the network increased three to four times. Then, center loss was added to the original loss function to improve the network's ability to distinguish different types of defects. Finally, a surface defect dataset containing 4655 images of 6 classes was established, and the proposed networks were trained on it. The proposed networks achieved 98.32 % accuracy with an average speed of 0.05s per image. Experimental results show that the slighter Faster R-CNN outperforms other steel strip surface defect detection methods in both accuracy and speed.