Demo-Net: A Low Complexity Convolutional Neural Network for Demosaicking Images
Mert Bektas, Zhao Gao, Eran A. Edirisinghe, A. Lluis-Gomez · 2022
This paper presents a novel Convolutional Neural Network (CNN) and an associated effective training approach that can be used for demosaicking images generated by different Color Filter Array (CFA) patterns, used in imaging sensors. The proposed CNN, Demo-Net, is a low complexity, auto-encoder based generalized CNN architecture, that can specifically take a CFA pattern as an additional input during training, thus creating a trained model for demosaicking images created by the specific CFA. The proposed Demo-Net allows one to create low complexity demosaicking systems that can be effectively deployed in consumer electronic devices with known sensor specifications.