Unsupervised demosaicking network using the recurrent renovation and the pixel-wise guidance
Jinyang Li, Jia Hao, Geng Tong, Shahid Karim, Xu Sun, Yiting Yu · Optics Letters · 2022
Demosaicking has recently been extensively studied and has achieved significant progress via deep learning. However, all the examples are trained in a supervised manner with the attendance of full-resolution polarization images, which has been compromised for practical applications. In this Letter, we propose to recover full-resolution images from a single mosaic image by combining the deep image prior with the polarization prior to capture the image-specific statistics and further guide the optimization. Specifically, we employ the pixel-wise weight on the intermediate outputs being generated by the recurrent strategy to self-supervise the learning, and the missing pixels can be iteratively and individually recovered. Experimental results on long-wave infrared (LWIR) polarization images demonstrate the effectiveness of the proposed method in terms of both quantitative measurement and visual quality.