A Real-time and High-efficiency Surface Defect Detection Method for Metal Sheets Based on Compact CNN

Abstract

Surface defect detection has received increased attention in the quality control of industrial products. It is an urge need to develop a real-time, high-efficiency defect detection algorithm on automatic detection equipment. The traditional image processing method based on connected domain does not meet the throughput requirement and cost huge efforts. In this paper, A novel detection algorithm based on compact convolutional neural network (CNN) is applied to confirm the defect's existence in the target region of an image. Experimental results show that our proposed compact CNN detection kernel finishes in 7ms per image and it achieves 8.57x speedup when compared to the traditional image processing method. Our compact CNN-based real-time processing pipeline also achieves 96.85% detection accuracy rate, which is more accurate and robust than human detection and traditional image processing algorithms. For the whole image processing pipeline, it achieves 1.79x throughput improvement in terms of image per second and higher efficiency in terms of out-of-pocket cost.