基于改进RRD-Net的低照度图像增强网络  

Low-Light Image Enhancement Network Based on Improved RRD-Net

作　　者：李谦一 张宏宇 房媛[1] LI Qianyi;ZHANG Hongyu;FANG Yuan(School of Information Science and Engineering Technology,Dalian Polytechnic University,Dalian116034,China)

机构地区：[1]大连工业大学信息科学与工程学院,辽宁大连116034

出　　处：《照明工程学报》2024年第5期61-67,共7页China Illuminating Engineering Journal

基　　金：辽宁省教育厅科研基金(JYTMS20230416);辽宁省自然基金规划项目(2022-BS-263);辽宁省“十四五”教育科学规划课题(JG21DB054);大连工业大学教学改革项目(JGLX2023028)。

摘　　要：为解决低照度环境下图像曝光度不足问题,本文提出了一种基于Retinex模型通过分解输入图像、以RRD-Net为先验知识的图像增强方法。对图像的反射图像进行自适应均衡化,调整原始图像生成噪声图像,结合照明分量生成灰度图。最后将图像进行整合,形成经过曝光修复的全新图像。该方法旨在改善低照度条件下图像的质量,提高可见度,增强物体细节,减少噪点,具有重要的实际作用。再通过自然图像质量评价(Natural Image Quality Assessment,NIQE)、峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)和结构相似性指数(Structural Similarity Index,SSIM)三种方法对增强后的图像进行客观评估。经评估,NIQE值为5.63,SSIM值为0.78,PSNR高达16.73。结果表明,本文算法相比于其他算法,图像增强效果显著。To address the issue of insufficient exposure in low-light environments,this paper proposes an image enhancement method based on the Retinex model,which decomposes the input image using RRD-Net as priori knowledge.Adaptive equalization is performed on the reflectance image,adjusting the original image to generate a noise image,and combining it with the illumination component to create a grayscale image.Finally,the images are integrated to form a newly exposed-repaired image.This method aims to improve the quality of images under low-light conditions,enhance visibility,highlight object details,and reduce noise,demonstrating significant practical value.The enhanced images are then objectively assessed using three methods:namely Natural Image Quality Evaluation(NIQE),Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity Index(SSIM).The NIQE value is assessed to be 5.63,the SSIM value is 0.78,and the PSNR reaches 16.73.The results indicate that the proposed algorithm significantly outperforms other algorithms in image enhancement.

关 键 词：图像增强 深度学习 图像去噪 低照度 改进Retinex 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]