Residual Convolutional Neural Networks Model For Image Denoising On Real Time

Abstract

This Nowadays, deep learning is one of the most used technique for image denoising until it outperforms so far, all other denoising methods. However, this method requires a lot of computing power, so it’s quite difficult to achieve real-time deep learning denoisers especially on edge devices like embedded systems and mobile phones. In this paper, we proposed a deep learning denoiser that works in real-time on a Raspberry Pi 3B+, 1GB of ram, to increase in real-time the incoming noisy video from a Raspberry Pi Camera frame per frame, where each frame is an RGB image if size 256x256. We used a residual denoiser that extracts the noise and enhance the quality of obtained images. In fact, the proposed architecture has a very small size that can fit easily on any edge device. Furthermore, many optimization techniques were applied on the denoiser so it can run faster on a very limited computing resource. Each denoised frame where uploaded directly to a Microsoft storage service.