Enhanced Faster Region Convolutional Neural Networks for Steel Surface Defect Detection

Abstract

Bar steel surface defects detection is very important to steel production and quality control. Many traditional computer vision methods have been applied to industrial defects detection, but they are usually environmentally sensitive and not robust enough. In this paper, a deep learning defects detection method based on Faster Region Convolutional Neural Networks (Faster R-CNN) is proposed. Firstly, to solve the problem of missed detection of a large number of small defects, we introduce Weighted Region of Interest (RoI) Pooling instead of RoI pooling, which eliminates the area misalignment caused by the two quantization processes in the latter, and the small defects detection rate is significantly improved. Secondly, considering that most of the defects are irregular in shape, we use deformable convolution in upper layers to adapt to various shapes by learning the positional offset in convolution. Thirdly, owing to the diversity of bar steel defects, multi-scale feature extraction network with Feature Pyramid Networks (FPN) is proposed to build feature pyramids. Finally, we propose Strict-Non-Maximum Suppression (Strict-NMS) algorithm to reduce overlapping bounding boxes as much as possible. Experiments on defect datasets in real industrial environments show that the detection rate of this method can reach 97%, which is much higher than state-of-the-art methods.