Demosaicking customized diffusion model for snapshot polarization imaging

Abstract

For snapshot polarization imaging, the color polarization demosaicking is essential to reconstruct full resolution from a mosaic array, which is the latest unsolved issue. Due to the mosaic array missing a large number of key pixels, existing one-step deep learning-based methods exhibit limited demosaicking performance. Hence, we make the first attempt to address the color polarization demosaicking task through the diffusion model, namely DCPM. Specifically, we extend the residual-based diffusion process to the task of color polarization demosaicking and improve the network architecture to accommodate full-resolution polarization images. Moreover, considering the polarization property of images, a customized loss function is proposed to assist in the diffusion model training. Extensive experiments on both synthetic and real-world benchmarks demonstrate the effectiveness of the proposed method. The source code will be available at https://github.com/JJGNB/DCPM.