In-Camera Raw Compression: A New Paradigm from Image Acquisition to Display
Zhihao Li, Haojie Liu, Lin Yang, Zhan Ma · 2020
We propose the NeuralCAM- a new camera processing paradigm where we directly compress the sensor raw Bayer data (RAW) and apply the image signal processor (ISP) at users’ client for on-demand rendering and processing. Both RAW compression and ISP processing are built upon the deep neural network (DNN)-based approaches that fully leverage the power of data-driven method to efficiently capture the underlying statistics for improved processing capacity. Experiments have shown 3 - 4 dB PSNR gains when compared with state-of-the-art image coders in RAW format, and have also revealed superior image quality in RGB space by mimicking the Professional ISPs that often couldn’t afforded by power-constrained mobile chips.