Color optimization and diffusion-based post-processing to obtain sharper images without compromising R-D performance in learned image compression

In the digital era, efficient storage and transmission of visual signals have become paramount due to the explosive growth in multimedia content. The need for advanced image compression methods is driven by increasing image resolutions and the limitations of traditional codecs in terms of flexibility and adaptability. In the first part of this thesis, we introduce a flexible method for coding color images in the YCrCb space, addressing the human visual system's greater sensitivity to the luma component over chroma components. We extend the variable-rate image coding approach to YCrCb images, enabling separate rate adjustments for luma and chroma components. By implementing image-adaptive luma-chroma bit allocation during inference, we can increase Y PSNR at the expense of slightly lower chroma PSNR, resulting in sharper images without introducing color artifacts. This strategy enhances image sharpness more effectively than optimizing for RGB PSNR alone. Our experimental results demonstrate that sharper images with better VMAF and Y PSNR can be obtained by optimizing models for YCrCb MSE compared to state-of-the-art models optimizing RGB MSE at the same bpp. In the second part, we explore the use of diffusion models for post-processing in wavelet-based image codecs. Diffusion models, a type of deep generative models, have shown great promise in various domains, including inverse problems in image processing. They are particularly effective at producing visually pleasing textures. By integrating a fixed, invertible transform with a learned entropy model and a diffusion-based post-processing module, we demonstrate enhanced visual quality without compromising the rate-distortion performance. Our experimental results show that sharper images with better perceptual quality and YCrCb PSNR can be obtained compared to state-of-the-art classic and learned codecs.