基于HSV与级联卷积神经网络的图像增强模型在低照度图像中的应用  

Application of Image Enhancement Model Based on HSV and Cascaded Convolutional Neural Network in Low Light Images

作　　者：刘晓伟[1] 刘迪 LIU Xiaowei;LIU Di(School of Mathematics and Physics,Bengbu University,Bengbu Anhui 233030)

机构地区：[1]蚌埠学院数理学院,安徽蚌埠233030

出　　处：《湖北理工学院学报》2024年第6期55-60,共6页Journal of Hubei Polytechnic University

基　　金：安徽省十大新兴产业项目(项目编号:2023sdxx097)。

摘　　要：为改善恶劣情况下低照度图像亮度不均匀、细节信息丢失等问题,将传统的红、绿、蓝(RGB)颜色空间转换到色调、饱和度、亮度(HSV)颜色空间,通过HSV颜色空间进行分量提取和图像增强。利用特征提取网络和纹理细化网络构建级联卷积神经网络,并与视觉感官色彩空间相结合,进行低照度图像的增强。采用正则化缓解模型的过拟合现象,并通过仿真实验对比验证了该模型在低照度图像上的增强效果,亮度、对比度、饱和度以及色调方面均表现出良好的增强效果。模型在RELLISUR数据集和Exclusive-Dark数据集上的峰值信噪比分别为24.32和26.70,结构相似度分别为0.97和0.96,对比度分别为9.12和8.89,在两个数据集上的运算时间分别为1.33 s和1.38 s。除运算时间方面较为落后外,其余指标均为最优值,并且结构相似度最接近于1,表明该模型能够有效还原图像细节信息,具有良好的科学性与实用性。To address image quality issues such as uneven brightness and loss of detail information in low light images under harsh conditions,this study converts the traditional red,green and blue(RGB)color space to the hue,saturation,and brightness(HSV)color space,and used the HSV color space for component extraction and image enhancement.A cascaded convolutional neural network using feature extraction network and texture refinement network was built,combined with visual sensory color space to enhance low light images.The study used regularization to alleviate the overfitting phenomenon of the model,and verified the enhancement effect of the model on low light images through simulation experiments.The proposed model showed good enhancement effects in brightness,contrast,saturation,and hue,and achieved peak signal-to-noise ratios of 24.32 and 26.70 on the RELLISUR dataset and Exclusive-Dark dataset,structural similarity of 0.97 and 0.96,and contrast of 9.12 and 8.89,respectively.The computation times on the two datasets are 1.33 s and 1.38 s,respectively.Except for the relatively backward computation time,all other indicators are the optimal values,and the structural similarity is closest to 1,indicating that the model can effectively restore image detail information and has good scientificity and practicality.

关 键 词：低照度图像增强 HSV颜色空间 卷积神经网络 计算机视觉 正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]