Remove glasses diffusion model an innovative conditioned of eye glasses removal with image diffusion model

Abstract

The presence of eyeglasses in facial images poses challenges for image processing, particularly in facial recognition. This paper introduces the remove glasses diffusion model (RGDM), a conditioned denoising diffusion probabilistic model (DDPM) designed for precise glasses removal. RGDM employs conditional modeling to focus on the glasses region while seamlessly restoring facial features. An eyes position accuracy mechanism, leveraging facial landmarks, ensures accurate eye restoration post-removal. Comprehensive evaluations on the CelebA dataset demonstrate RGDM’s superior performance, achieving the lowest Fréchet inception distance (FID) of 27.09 and learned perceptual image patch similarity (LPIPS) of 0.299, outperforming state-of-the-art methods such as 3D synthetic, cycleconsistent generative adversarial network (CycleGAN), and eyeglasses removal generative adversarial network (ERGAN). These results highlight the model’s effectiveness in producing natural and high-fidelity facial reconstructions. This work advances glasses removal technology and underscores the significance of conditional models in image processing. The proposed approach has practical implications for facial recognition and image enhancement, paving the way for more accurate and robust real-world applications.