Lightweight CNN-Based Low-Light-Image Enhancement System on FPGA Platform

Wei Wang, Xin Xu

Abstract: Deep learning methods have made great advances in low-level vision tasks, such as image enhancement and target detection. However, these methods cannot be executed on mobile on-chip platforms for their high memory occupancy and multiple hundreds of Floating Point Operation per Second (FLOP/s). In this paper, a lightweight convolutional neural network (CNN) for low light image enhancement task is proposed with less than 1 M parameter size and solving the problem of hyper parameter setting in enhancement models. Then, a pseudo-symmetry quantization method to introduced for image enhancement model compression, since recent quantization approaches aim at high compression ratio and are not suitable for image enhancement task. At last, this CNN-based low light enhancement method is deployed on a customized Xilinx Inc. FPGA platform with hardware accelerator, which is designed to speed up the multi threads data transmission. The experiments have proved that our method gives visual and objective convincing results and potential to be executed for real-world applications.