Research And Hardware Implementation Of Low-Light Image Enhancement Algorithm Based On Deep Learning

In low light,uneven light and backlight environments,captured images suffer from image degradation such as insufficient brightness,missing detail information and noise interference,which not only degrade people’s visual perception,but also seriously affect the performance of computer vision task systems such as target recognition and tracking and intelligent driving assistance.The purpose of low-light image enhancement is to process low-quality images with insufficient illumination and restore the brightness,contrast,and detailed texture information of the images,so that the image quality can be significantly improved.In this paper,we address the problems of lack of contrasted normal illumination images and the excessive amount of network parameters encountered in the research of low-light image enhancement algorithms,and improve and optimize the low-light image enhancement algorithm based on self-regularization.The network structure is based on the U-Net network.In the network structure of this paper,the multi-scale channel attention feature fusion is applied to the feature maps of the corresponding layers of the encoder and decoder parts of the backbone network,and the fusion of the bottom features and the top features enables the network to retain more highresolution information,which can better recover the detailed information contained in the image and improve the quality of the recovered image.The improved algorithm is trained and tested on the SCIE dataset,and the experimental results show that the number of parameters of this algorithm is only 12.6MB compared with that of the self-regularization algorithm,and the recovered images have improved by 0.19 and 0.004 in peak signal-to-noise ratio and structural similarity,respectively.The effectiveness of the improved algorithm is demonstrated by objective comparison of the evaluation indexes of peak signal-to-noise ratio and structural similarity.In this paper,the hardware of the improved channel attention feature fusion image enhancement network is deployed based on ZYNQ-7020 development platform.Firstly,INT8 quantization of the weight and Bias parameters of the trained network model is carried out,and then the hardware and software division is carried out according to the network structure and operation process.The hardware design completes the operation of convolution,pooling,bilinear interpolation,activation function and other modules on the PL side.The software design schedules and controls the operation and data interaction on the PL side through the register configuration design on the PS side,and completes the operation of the multi-scale channel attention feature fusion module.Finally,the verification test was carried out on ZYNQ-7020 development platform.When the system clock frequency was 100MHz,the overall reasoning time of the algorithm was only 1.57s,and the power consumption was 4.85W.