Multi-Scale Feature Pair Based R-CNN Method for Defect Detection

Abstract

The traditional defect detection algorithms based on image registration, image contrast and other image processing algorithms are only limited to a single defect. Though deep-learning-based object detection algorithms can be used to detect a variety of different defects, the state-of-the-art deep-learning-based object detection algorithms still have low detection accuracy on small size defects. Basing on Cascade R-CNN in this paper, a new multi-scale feature extraction method-the Multi-Scale Feature Pair-is proposed and is used to establish a defect detection model for metal can products of an enterprise. Experimental results show that the accuracy (AP@0.5) of our improved model is 6.1% higher than Cascade R-CNN.