CameraNet: A Two-Stage Framework for Effective Camera ISP Learning

Traditional image signal processing (ISP) pipeline consists of a set cascaded modules onboard camera to reconstruct high-quality sRGB from the sensor raw data. Recently, some methods have been proposed learn convolutional neural network (CNN) improve performance traditional ISP. However, in these works usually CNN is directly trained accomplish ISP tasks without considering much correlation among different components an As result, quality reconstructed images barely satisfactory challenging scenarios such as low-light imaging. In this paper, we firstly analyze ISP, and categorize them into two weakly correlated groups: restoration enhancement. Then design two-stage network, called CameraNet, progressively groups tasks. each stage, ground truth specified supervise subnetwork learning, subnetworks are jointly fine-tuned produce final output. Experiments on three benchmark datasets show that CameraNet achieves consistently compelling reconstruction outperforms recently learning methods.