MSC-DNet: An efficient detector with multi-scale context for defect detection on strip steel surface

Abstract

The strip steel has been widely used in the manufacturing industry. Defects on the surface are main factors to determine the quality of strip steel. Due to the various shapes of the defects and background interference, the CNN-based algorithm cannot give full play to its best performance. In this paper, a defect detection module, named detection network with multiscale context (MSC-DNet), is proposed to localize the precise position of defect and classify the specific category of surface defects. In MSC-DNet, a parallel architecture of dilated convolution (PADC) with different dilation rate is built up to capture the multi-scale context information containing multiscale defects. Furthermore, a feature enhancement and selection module (FESM) is proposed to enhance the single-scale features and select the multi-scale features for reducing the confusing information. During the training, the auxiliary image-level supervision (AIS) is adopted to speed up the convergence and to enhance the feature discrimination of the target defects. The experiment results show that the proposed MSC-DNet reaches the accuracy of 79.4% mAP and 14.1 FPS on NEU-DET dataset, and 71.6% mAP on GC10-DET dataset among all the benchmark methods, which satisfies the quasi-real-time requirement in multiscale defect detection task.