Improved Method Based on Faster R-CNN Network Optimization for Small Target Surface Defects Detection of Aluminum Profile

Abstract

Aiming at the problems of missed detection and false detection of small target defects on the surface of aluminum profiles, an improved Faster R-CNN deep learning network is proposed to detect paint bubbles and dirty spots. Firstly, create a small target data set for these two types of defects, redesigning the anchor ratio, and then use ROI-Align instead of ROI-Pooling to obtain more accurate defects location information, and finally use Soft-NMS algorithm to replace NMS Algorithm to eliminate redundant prediction frames, which improves the detection accuracy of small target defects. The experiment shows that the improved network has an AP value of 64.06% for paint bubbles defects, which is 12.99 % higher than the original Faster R-CNN network, and the AP value of dirty spots defects is up to 82.98%, which is 6.59% higher than the original Faster R-CNN network.