首页|基于照度图引导的低照度图像增强网络

基于照度图引导的低照度图像增强网络

扫码查看
原文链接
  • 万方数据
  • 维普
在低照度环境下采集的图像，由于光照的不均匀性，存在能见度差、对比度低和颜色失真等问题。现有的大多数低照度图像增强方法存在过增强或欠增强的现象，影响视觉感知和后续目标检测任务。针对上述问题，提出一种基于照度图引导的低照度图像增强网络。首先根据低照度图像的灰度分布特点构造对应的照度图，度量低照度图像不同区域块的明暗程度;然后利用照度图作为网络增强的引导图，与低照度图像一起送入图像增强网络来获得增强后的图像。为了解决训练数据不足的问题，提出一种基于内循环和概率旋转的数据增强方法来扩充训练数据样本的数量和多样性;同时，针对目前图像增强方法中普遍存在照度不均匀的问题，基于直方图匹配的思想构建一种直方图损失函数，约束并指导网络的训练。在合成数据集 LOL 和真实图像上的实验结果表明，所提网络在低照度图像增强方面获得了更好的主观视觉效果;与经典的RetinexNet方法相比，所提方法在PSNR和SSIM客观定量指标上分别提高了 7。905 dB和 0。328;该网络对后续目标检测任务的检测率可提高 10。17%～17。19%。
Low-Light Image Enhancement Network Guided by Illuminance Map
Images captured in low-light environment suffer from poor visibility,low contrast and color dis-tortion due to the uneven illumination.Most of the existing low-light image enhancement methods have problems of over-or under-enhancement,which affects visual perception and subsequent object detection tasks.To address these problems,this paper proposed a low-light image enhancement network based on il-lumination map guidance.First,according to the grayscale distribution characteristics of the low-light im-ages,the corresponding illumination map is constructed to measure the brightness and darkness of different areas of the low-light image;then,the illumination map is regarded as a guidance map and fed into the image enhancement network together with the low-light image to obtain the enhanced image.In addition,in order to solve the problem of insufficient training data,a data enhancement method based on inner loop and prob-ability rotation is proposed to expand the number and diversity of training data samples;simultaneously,a histogram loss function is designed based on the idea of histogram matching to constrain and guide the training of the network to overcome the problem of uneven illumination in current image enhancement methods.Experimental results on synthetic dataset LOL and real images demonstrate that the proposed net-work achieves better subjective visual effects in low-light image enhancement.Compared with the classical RetinexNet method,the proposed method improves the objective quantitative indexes of PSNR and SSIM by 7.905 dB and 0.328,respectively;moreover,the detection rate of the proposed network for subsequent object detection tasks can be improved by 10.17%to 17.19%.

low-light image enhancementguidance of illuminance maphistogram loss functionprobability rot-ation enhancementobject detection

黄淑英、黎为、杨勇、万伟国、赖厚增

展开 >

天津工业大学软件学院 天津 300387

江西财经大学软件与物联网工程学院 南昌 330032

天津工业大学计算机科学与技术学院 天津 300387

低照度图像增强 照度图引导 直方图损失函数 概率旋转增强 目标检测

国家自然科学基金国家自然科学基金江西省自然科学基金江西省自然科学基金

618620306207221820192ACB2000220192ACBL21008

2024

10.3724/SP.J.1089.2024.19779

计算机辅助设计与图形学学报
中国计算机学会

计算机辅助设计与图形学学报

CSTPCD北大核心
影响因子：0.892
ISSN：1003-9775
年,卷(期)：2024.36(1)
  • 27