High Quality Coal Foreign Object Image Generation Method Based on StyleGAN-DSAD.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Coal mine safety and efficiency rely on accurate foreign object detection.
• Limited foreign object image datasets pose a significant challenge for deep learning models.
• Data augmentation is crucial for improving the robustness of foreign object detection systems.

Purpose of the Study:

• To propose a novel method for generating high-quality coal foreign object images using an improved StyleGAN.
• To address the challenge of scarce datasets in coal foreign object detection.
• To enhance the performance of foreign object detection models through effective data augmentation.

Main Methods:

• An improved StyleGAN architecture incorporating a dual self-attention module in the generator.
• Integration of depthwise separable convolution in the discriminator for increased efficiency.
• Comparative analysis of image generation quality and diversity against classical GANs and original StyleGAN.
• Evaluation of the proposed method's effectiveness in foreign object detection tasks.

Main Results:

• The improved StyleGAN significantly enhances the quality and diversity of generated images, evidenced by a 2.52 IS improvement and 5.80 FID decrease.
• Model complexity is reduced, with parameters and training time decreased to 44.6% and 58.8% of the original StyleGAN.
• The proposed image generation method outperforms traditional augmentation techniques in foreign object detection.
• Detection performance improved by 5.8% in AP_box and 4.5% in AP_mask.

Conclusions:

• The improved StyleGAN effectively generates high-quality coal foreign object images, mitigating dataset scarcity.
• The enhanced model offers superior image generation quality and diversity with reduced computational complexity.
• This data augmentation approach significantly boosts the performance of coal foreign object detection systems.