Smart embedded system based on demosaicking for enhancement of surveillance systems

Abstract

Demosaicking and denoising are essential elements in digital photography pipelines. The use of convolutional neural networks (CNN)-based image demosaicking and denoising methods has been very successful. However, still there is a room for improvement in the network performance in terms of efficiency and accuracy. The main challenge that remains to be addressed is to guarantee the visual quality of reconstructed images, particularly in the presence of noise. To address these challenges, this paper introduces a novel demosaicking and denoising conjunct strategy using deep adaptive residual learning. The proposed framework has three stages. Initially, zero padding is performed to increase processing speed and preserve the edges of the image. In the second phase, we perform demosaicking using interpolation in order to find missing values using information about neighboring pixels. Finally, the reconstructed image is created using the original image. To evaluate the feasibility of the proposed scheme, we used Pytorch and Google Colab with 400 images for training and 100 images for validation The outcomes show that the proposed scheme beats cutting edge joint demosaicking and denoising schemes regarding both structural similarity index metrics (SSIM) and peak signal-to-noise ratio (PSNR) and basic similitude record measurements (SSIM).