An Energy-Efficient Image Denoising Accelerator with Depth-wise Separable Convolution and Fused-Layer Architecture

Abstract

Image denoising is an important low-level vision task, which has been widely studied to reduce the noise in images. Denoising methods based on deep learning have achieved great performance improvement. However, the huge computation requirements of these methods prevent their application in practical scenarios. Moreover, the hardware accelerator of deep learning denoising methods has rarely been studied. Therefore, we optimize DnCNN for additive white Gaussian noise (AWGN) to obtain the hardware-friendly Light-DnCNN and design an energy-efficient denoising accelerator based on Light-DnCNN. The accelerator has a denoising frame rate of 19.9 FPS with 3.52 W. Its energy efficiency is 5 times and 1319 times higher than that of Titan X GPU and Intel i7-9700 CPU respectively.