RISPNet: A Network for Reversed Image Signal Processing

Xiaoyi Dong, Yu Zhu, Chenghua Li, Peisong Wang, Jian Cheng

Abstract: RAW data is considered to be valuable owing to its beneficial properties for downstream tasks, such as denoising and HDR. Usually, RAW data gets rendered to RGB images by the in-camera image signal processor (ISP) and shall not be saved in most cases. In addition, RGB images produced by non-linear operations in ISP would necessitate the execution of other image processing tasks. To overcome this problem and acquire RAW data again, we propose a new reversed ISP network (RISPNet) to achieve efficient RGB to RAW image conversion. Our main proposal is a novel encoder-decoder network with a third-order attention module. Since Attention facilitates the complex trade-off between restoring spatial details and high-level contextual information, the accuracy of recovering high dynamic range RAW data is greatly improved, and the whole recovery process becomes a more manageable step. Benefiting from the design of the attention module, our RISPNet has a remarkable ability to recover RAW images. According to the AIM 2022 Reversed ISP Challenge, RISPNet achieved third place on both Track P20 and Track S7 test sets.