Improving image visual quality of petroglyph images using ESRGAN (enhanced super-resolution generative adversarial networks)

The traditional human-based method of petroglyph image detection and interpretation requires extensive resources and personal judgment. The combination of automatic image enhancement through modern machine learning algorith ms brings a revolutionary solution to archaeologists and historians. The research applies [2] Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) [2] to restore petroglyph images which suffer from deterioration caused by erosion and vandalism and environmental damage. Our research introduces three main contributions to petroglyph image processing: (i) ESRGAN adaptation for petroglyph-specific image restoration needs and (ii) a two-stage training approach that optimizes both perceptual quality and pixel precision and (iii) a comprehensive evaluation against different baseline methods. The evaluation of our method uses RMSE and Perceptual Index (PI) metrics on DIV2K and an Indian petroglyph collection with future plans to add SSIM and PSNR measurements. The results show ESRGAN produces the most realistic results (best PI score) while uncovering concealed petroglyph details at the expense of slightly elevated RMSE values relative to other optimized techniques. The research supports the implementation of ESRGAN in digital archaeology because it provides a method for dependable and expandable cultural heritage image improvement.