A Defect Detection Network Based on An Improved Receptive Field Block

作者：Yangsai Wang, Yulong Zhuang, Yongqi Han · 年份：2021 · DOI：10.1109/mlise54096.2021.00024 · 研究领域：Industrial Vision Systems and Defect Detection、Advanced Neural Network Applications、Infrastructure Maintenance and Monitoring

The traditional defect detection algorithms such as Edge Examination have already shown promising results for simple background industrial product images defect detection. they are insufficient for complex texture industrial product images detection capabilities because the background information of complex texture images is similar to the type of defects. To achieve the classification and location of defects in complex texture industrial product images, in this paper, based on Receptive Field Block Net (RFB Net), we extend the RFB module and couple the lower layer with the higher layer for defect detection. Experimental results show that the accuracy (AP@0.5) of our improved model is 1.7% higher than RFB Net on complex texture industrial product images, 2.7% higher than RFB Net on simple background industrial product images.