融合注意力机制和高斯滤波的Retinex图像增强算法  

Retinex Image Enhancement Algorithm Incorporating Attention Mechanism and Gaussian Filtering

作　　者：张亚蒙 张宝菊[1,2] 张博[1,2] 张翠萍[1,2] 褚博华 姜凯 薛梦奇 孙艺 隋晓楚 ZHANG Yameng;ZHANG Baoju;ZHANG Bo;ZHANG Cuiping;CHU Bohua;JIANG Kai;XUE Mengqi;SUN Yi;SUI Xiaochu(College of Electronic and Communication Engineering,Tianjin Normal University,Tianjin 300387;Tianjin Key Laboratory of Wireless Mobile Communication and Power Transmission,Tianjin Normal University,Tianjin 300387)

机构地区：[1]天津师范大学电子与通信工程学院,天津300387 [2]天津师范大学天津市无线移动通信与电力传输重点实验室,天津300387

出　　处：《软件》2025年第1期20-23,共4页Software

基　　金：天津市普通高等学校本科教学改革与质量建设研究计划A231006507号项目;教育部中国高校产学研创新基金2022BL084号项目;天津市教委科研计划项目(2024KJ061);工业和信息化部教育与考试中心2024年度研究课题、工信部项目。

摘　　要：本文在Retinex-Net网络的基础上,提出了一种新的低光照图像增强网络,以解决Retinex-Net网络中反射分量噪声大,光照分量亮度低且特征提取不足等问题,有效增强了低光图像的质量。在分解网络中使用Retinex-Net网络作为基本模型,将普通卷积与空洞卷积结合,进行多尺度特征抽取,从而获得更为丰富的细节信息;在增强网络中,采用多层注意力机制模块,实现对图像中细节部分和光照部分的增强。最后,通过结合去除噪声的反射分量和增强后的光照分量,生成一幅标准的正常光照图像。实验结果表明,本文所提出的算法能够有效地增强低光照图像的亮度和对比度,同时更好地保留图像中细节信息,从而提高图像的视觉效果。Based on the Retinex-Net network,this paper proposes a new low-light image enhancement network to address the problems of high noise in the reflection component,low brightness in the illumination component,and insufficient feature extraction in the Retinex-Net network,effectively enhancing the quality of low-light images.Firstly,the decomposition network uses the Retinex-Net network as the basic model,combining ordinary convolution with atrous convolution to extract multi-scale features,thus obtaining more detailed information;in the enhancement network,a multi-layer attention mechanism module is adopted to enhance the details and illumination of the image.Finally,by combining the noise-reduced reflectance component with the enhanced illumination component,a standard normal illumination image is generated.Experimental results demonstrate that the proposed algorithm enhances the brightness and contrast of low-light images,while preserving details and improving the overall visual quality.

关 键 词：RETINEX算法 注意力机制 低照度图像 图像增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]