A Lightweight Convolutional Neural Network for Camera ISP

Mobile cameras are ubiquitous in our daily life. The widely used sRGB (standard RGB) images are generated from the Raw data by a series of image signal processing (ISP) operations, such as demos acing, white balance, tone mapping, etc. In this paper, we realize the camera imaging process based on a convolutional neural network, that is, using an end-to-end convolutional neural network instead of the traditional image processing algorithm. Specifically, we first propose a pyramid U - N et for camera ISP and then optimize it to a light weight network. In addition, we extend the trained model to different mobiles by using a few images to finetune the pre-trained network. Experimental results demonstrate that the proposed network achieves robust ISP result with the fewest parameters and inference time.