Application of improved Faster R-CNN algorithm in digital printing fabric defect detection

Abstract

To address the problem of unbalanced distribution of class samples in the self-built data set in digital printed fabrics defect detection, which leads to the low detection accuracy, an improved Faster R-CNN defect detection algorithm was proposed by introducing adaptive class suppression loss (ACSL). Firstly, according to the characteristics of complex defect background and difficult target detection of digital printed fabrics, a defect detection network structure based on Faster R-CNN was constructed. Then, in order to maintain the relative balance of positive activation loss and negative activation loss of different class positions, an ACSL module was introduced into the basic network, which was used to adjust the weight coefficient of the classification loss of different class positions, and reduce the effect of sample imbalance on detection accuracy. The experimental results show that the proposed algorithm achieves the average precision of 0.60 on the COCO standard, which is 0.02 higher than that of Faster R-CNN, and that the proposed method can effectively solve the problem of digital printing defect detection with unbalanced class samples.

Keywords

• digital printing
• defect detection
• imbalanced class sample
• adaptive class suppression loss (acsl)
• faster r-cnn