An End-to-End Real-World Camera Imaging Pipeline

Demosaicing Denoising Tone Mapping Encoder (JPEG or BPG or NN) Decoder (JPEG or BPG or NN) (a) Traditional: Module Separation Design.(b) Learning-based: Module Separation Design.ISPNet (c) Learning-based: Partial Joint Module Design.Encoder (d) Ours: Fully End-to-End RealCamNet.Decoder CIMC CADR Figure 1: (a)Traditional Design: Implements each step of Image Signal Processing (ISP) separately using conventional algorithms.(b) Learning-based Separated Design: Employs neural networks to individually address each phase of the ISP process.(c) Learningbased Partial Joint Design: Develop an ISPNet to unify operations such as demosaicing, and tone mapping.(d) RealCamNet:We propose an end-to-end camera imaging framework that categorizes ISP operations into coordinate-independent and coordinatedependent groups.CIMC and CADR are designed to perform tasks like demosaicing and image feature compression and to restore coordinate-dependent image distortions (e.g., vignetting, dark shadows), respectively.