摘要
针对不同光照下交通标志图像检测与识别困难的问题,提出一种基于Retinex-Gamma的光照图像增强算法,该算法与Mask R-CNN相结合,称为Retinex-Gamma-Mask R-CNN算法.首先,基于光照反射成像模型将图像RGB空间转换为HSV空间,对V通道进行多尺度高斯滤波处理获得光照分量,利用光照分量提取反射分量,并对反射分量进行线性拉升优化;其次,利用光照分量的分布特征进行二维Gamma函数调整,并获得优化后的亮度分量;最后,利用混合空间增强法获得增强后的V通道,重新构造图像.实验采用的ZCTSDB数据集共有15724幅图像,包含不同光照的驾驶环境.实验结果表明,与标准Mask R-CNN相比,Retinex-Gamma-Mask R-CNN算法对交通标志的目标检测的均值平均精度提升了0.161%,对交通标志的实例分割的均值平均精度提升了0.363%.
In order to deal with the difficulty of traffic sign image detection and recognition under different illumination condition,an illumination image enhancement algorithm based on Retinex-Gamma is proposed.The algorithm is combined with Mask R-CNN,which is called Retinex-Gamma Mask R-CNN algorithm.Firstly,based on the illumination reflection imaging model,the image RGB space is transformed into HSV space,the V channel is processed by multi-scale Gaussian filtering to obtain the illumination component,the illumination component is used to extract the reflection component,and the reflection component is linearly optimized.Secondly,the two-dimensional Gamma function is adjusted by using the distribution characteris-tics of illumination components,and the optimized brightness components are obtained.Finally,the en-hanced V channel is obtained by using the mixed space enhancement method to reconstruct the image.The ZCTSDB dataset used for the experiment has a total of 15724 images and contains driving environments with different lighting.The experimental results show that compared with the standard Mask R-CNN,the average accuracy of Retinex-Gamma-Mask R-CNN algorithm for target detection of traffic signs is im-proved by 0.161%,and the average accuracy of instance segmentation of traffic signs is improved by 0.363%.
作者
项新建
姚佳娜
黄炳强
杨松
武晓莉
Xiang Xinjian;Yao Jiana;Huang Bingqiang;Yang Song;Wu Xiaoli(School of Automation and Electrical Engineering,Zhejiang University of Science and Technology,Hangzhou 310023;Intelligent Transportation Research Institute,Zhejiang Scientific Research Institute of Transport,Hangzhou 310009)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2023年第2期293-302,共10页
Journal of Computer-Aided Design & Computer Graphics
基金
浙江省公益技术研究计划项目(LGG19F030005).
