Image Inpainting on Archeological Dataset Using UNet Architecture on Embedded Platform

Image inpainting is a widely used technique for reconstructing damaged or missing portions of images, with applications in image editing, object removal, and image restoration. This paper proposes a novel approach to enhance the quality of image inpainting specifically in archaeological datasets. Our method focuses on enhancing the UNet architecture by incorporating the MS-SSIM loss function, which effectively denoises the output. We conducted experiments using datasets from archaeological sites, which often contain photographs with missing or damaged areas due to the passage of time or excavation procedures. Our results demonstrate that the proposed approach significantly improves the SSIM score, increasing it from 0.9026 to 0.9590. Additionally, we employed post-training quantization, resulting in a remarkable 77.1% reduction in the model size, making it more suitable for deployment in mobile applications. The suggested approach enhances image restoration and analysis for archaeological data, providing academics and archaeologists with a valuable tool for their work. Furthermore, by integrating the model into mobile apps, we aim to expand its usability and accessibility, empowering users to examine and restore their own archaeological images effectively.