An energy-efficient dehazing neural network accelerator based on E\(^2\)AOD-Net

Turbid media such as fog and haze seriously affects the quality of imaging for systems such as urban surveillance and satellite remote sensing. Image dehazing has become a research hotspot in the field of computer vision. Neural-network-based image dehazing has the potential of high performance, but requires high computational power and storage space, making it costly to deploy in a system with limited hardware resources, especially for edge computing systems. In this paper, we propose an energy-efficient dehazing neural network named E \(^2\) $_{}^2$ AOD-Net, which is pruned (generalization performance rises), quantized, pipelined, and parallelized over AOD-Net. We implement E \(^2\) $_{}^2$ AOD-Net on FPGA platform, achieving a lightweight dehazing hardware accelerator that realizes real-time dehazing. The experimental results show that the frame rate of E \(^2\) $_{}^2$ AOD-Net hardware accelerator inference reaches 38.3 FPS, while consuming the power of 2.491 watts. The VMAF index is improved by 89.14%. The energy efficiency is 42.61 GOPS/w.