A novel MAS-GAN-based data synthesis method for object surface defect detection

Surface defect detection in industrial processes is an essential step in production. The use of surface defect detection technology is of great significance for improving product quality and increasing production efficiency. However, the number of surface defect samples collected in the industrial process is limited, making training deep learning-based object detection models challenging. This paper proposes a multi-scale progressive generative adversarial network (MAS-GAN) that combines non-leaking data augmentation and self-attention mechanisms to solve this problem. The model uses an asymptotic growth strategy to synthesize multi-scale surface defect images and uses a non-leaking data augmentation method to deal with the degradation of the performance of the generative model in the case of insufficient samples. The self-attention mechanism further optimizes the generative adversarial network to make the details of high-resolution images more perfect. Using MAS-GAN to synthesize surface defect images to assist in training a deep learning-based object detection algorithm, both the training convergence speed of the surface defect detection model and the detection accuracy is improved. The experimental results on different datasets show the effectiveness of the proposed data synthesis method in the detection of surface defects of objects.

Automated summary

This paper proposes a multi-scale progressive generative adversarial network (MAS-GAN) to address the issue of limited surface defect samples in industrial processes. The model combines non-leaking data augmentation and self-attention mechanisms to synthesize multi-scale surface defect images and improve the performance of the generative model. The self-attention mechanism further enhances the details of high-resolution images. Experimental results demonstrate the effectiveness of the proposed data synthesis method in the detection of surface defects.