FFDDNet: Flexible Focused Defect Detection Network

Abstract

It is an important research area to optimize object detection methodology for improving the performance of steel surface defect detection. However, existing research is suffering from the problem of non-rigid deformation, background noise interference, and low detection efficiency. To address this problem, we propose an end-to-end flexible focused defect detection network (FFDDNet) in this article. First, we propose a restricted deformable convolutional network (RDCN) to efficiently extract the defect features, which makes the receptive field more fit the pixel area of the target image within the limited range and avoid excessive movement of the sampling points by designing a restricted function. Second, we propose the global vertical and horizontal attention (GVHA) to effectively reduce the interference of irrelevant background information, which consists of nested vertical and horizontal attention mechanism. Finally, we combine RDCN and GVHA to form a new convolution operator global RDCN (GRDCN) and apply to the information calculation of the new detection network FFDDNet to achieve accurate and efficient detection. Multiple experimental results on the NEU-DET dataset and GC10-DET dataset show that the mean average precision (mAP) of our proposed model increases by 8.5% and 3.6%, respectively compared with baseline and maintains a fast detection speed. It means FFDDNet effectively improves the efficiency of defect detection on surfaces. The source codes are available at: https://github.com/LinZeyu12/FFDDNet.