Mosaic Reduction Model Using Local Thermal Diffusion Equation

In this paper, we propose Mosaic reduction model using local thermal diffusion equation(MRLTDE) based on local heat diffusion equation and the U-net architecture. Due to the inherent problems such as unsmooth mapping, the traditional GAN-based method cannot achieve the best image restoration effect, which is reflected in the high differences in structure, brightness, contrast, and other aspects of the restored image, as well as high noise, blur distortion, color deviation and other issues. Therefore, we use the diffusion model for image demosaicing tasks, we improve the thermal diffusion model, and create a new local diffusion method to create image sequences with small changes between adjacent frames, so as to go further than previous schemes in image generation effect. Experiments of our scheme on CelebA datasets show that our model outperforms the existing GAN-based methods in terms of detail recovery, evaluated by two indicators: PSNR(Focus on pixel-level accuracy) and SSIM (focuses on the structural similarity and visual perception of images).In our experimental results, this model outperforms the base model on the above two indexes, in addition to well-defined metrics, even by ordinary people’s visual standards, our generation results are better than base. Detailed codes and experimental results are available at https://github.com/lhyf10561/MRLTDE.