On-line surface defects detection of Engine blocks based on Improved CNN+GAN Networks

Abstract

In this paper, a machine vision system for surface defect detection on engine blocks in an engine manufacturing line is proposed. This system employs deep learning algorithms to address the issues of high defect omission rate, high false alarm rate, and slow detection speed in the current inspection process. Specifically, an improved pix2pix GAN model is proposed by introducing a ResNet network to improve the feature expression capability and generate images with lower Frechet Inception Distance (FID) score. Furthermore, the generated images are used to augment the training dataset and combined with the YOLOv5 CNN network to enhance the training effect on small sample data. This method is compared with Faster-RCNN and SSD detection networks to validate its superiority in defect recognition performance. Experiments show that the proposed system achieves higher accuracy and robustness, with an average recognition accuracy of 98.4% and a detection time less than 0.5 seconds per image. The system has been applied in the actual production line and achieved good performance, demonstrating its feasibility and potential for practical use in the manufacturing industry.