Generate DSLR-Like Image With Global Information and Prior Guided ISP

Mobile to DSLR ISP frameworks aim to mitigate the quality gap between Mobile and DSLR captured images using a learnable pipeline. In this paper, we further mimic DSLR images from the perspective of fidelity and perception with a two-stage ISP strategy. Brightness and color distribution of DSLR images are highly correlated with shooting conditions, hardware settings, and the build-in ISP system. Existing cross-device ISP works mainly focus on image to image transfer from Mobile Raw Patch to DSLR RGB Patch and ignore the global and device-related factors underneath. In this paper, we start with patch mapping and inject global feature through DSLR reconstruction, which enable the model to generate ISP results with high fidelity. After obtaining the initial results, we further refine the RGB images with a DSLR prior guided model and generate the final cross-ISP output through feature matching. The refinement aims to further boost the perceptual quality of the ISP. Experiments show that our method performs favorably against state-of-the-arts on ISPW dataset.