Project DetailsThe traditional image signal processing (ISP) pipeline involves a series of sequential image processing modules within a camera that transform raw sensor data into a high-quality sRGB image. Recently, methods have been developed to enhance traditional ISP performance using convolutional neural networks (CNNs). However, these approaches typically train a CNN to handle ISP tasks without adequately considering the interconnections among different ISP components. Consequently, the image quality in challenging scenarios, like low-light conditions, remains subpar.
This work first examines the relationships between various ISP tasks and categorizes them into two loosely related groups: restoration and enhancement. We then introduce a two-stage network, CameraNet, designed to sequentially learn these two groups of ISP tasks. Each stage is supervised with Ground-truth data, and the two subnetworks are fine-tuned together to produce the final image. Experiments on SID benchmark dataset demonstrate that CameraNet (two stage model) consistently delivers superior reconstruction quality, outperforming one stage model in terms of visual examination. Then, as part of the follow-up project we compare the performance of the CameraNet model with other models and test its performance and efficiency in depth using different loss functions and additional metrices, on the same dataset. The experiments that were done as part of the second work also shows the effectiveness of Camera-Net especially visually.
