Detection of Steel Surface Defect Based on Faster R-CNN and FPN

Abstract

With the development of economy, steel has been widely used by human beings in various fields of social life, for example, military products, vehicles, aerospace, high-rise buildings and daily necessities. Therefore, it is very important to get high quality steel in industrial production. Detection of steel surface defect is a vital process in steel quality inspection. Traditional method uses artificial feature to detect steel surface defect, which costs lots of time to get features, and it is not robust in new environment. Therefore, this paper proposes an efficient and accurate method to solve the problem. In this paper, we adopt the faster R-CNN algorithm as baseline to train our own model. The feature pyramid network (FPN) is added to the original network of faster R-CNN, so that the network can combine high-level feature information and low-level feature information, furthermore, we replace the region-of-interest (RoI) pooling module with RoI align to reduce quantization error, which helps the mean average precision (mAP) increase about 0.8%. We use cycle GAN to do data enhancement, which helps our network easily to converge. Another problem is that it is quite common to get extreme aspect ratio in steel surface samples, which makes the detection more difficult, we introduce a method called multi-layer RoI align to solve this problem, which makes the mAP increase about 3.2%. According to our experiment results, the proposed method has a quite good performance on detection of steel surface defect.