ParamISP is designed to faithfully reflect the real-world ISP operations that change based on camera parameters. It consists of a pair of forward (RAW-to-sRGB) and inverse (sRGB-to-RAW) ISP networks that incorporate a novel neural network module called ParamNet. ParamNet extracts a feature vector from the camera parameters (exposure time, sensitivity, aperture size, and focal length) and feeds it to the ISP subnetworks to control their behaviors.
To effectively learn ISP operations for varying camera parameters, ParamISP employs a non-linear equalization scheme to adjust the scales of the camera parameters and a random-dropout-based learning strategy. The ISP networks consist of four subnetworks: CanoNet, LocalNet, GlobalNet, and ParamNet. CanoNet performs canonical ISP operations like demosaicing, white balance, and color space conversion. LocalNet and GlobalNet learn local and global ISP operations, respectively, with ParamNet controlling their behavior based on the camera parameters.
Extensive experiments demonstrate that ParamISP achieves superior RAW and sRGB reconstruction results compared to previous methods. It also shows robust applicability to various applications such as deblurring dataset synthesis, raw deblurring, HDR reconstruction, and camera-to-camera transfer.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Woohyeok Kim... at arxiv.org 04-16-2024
https://arxiv.org/pdf/2312.13313.pdfDeeper Inquiries