基于对象语义分割的改进颜色恒常性算法  

Improved Color Constancy Algorithm Based on Object Semantic Segmentation

作　　者：徐亮山 张霞 范珮 XU Liang-shan;ZHANG Xia;FAN Pei(Research Center of Graphic Communication,Printing and Packaging,Wuhan University,Wuhan 430079,China;School of Civil Engineering,Wuhan University,Wuhan 430079,China;Shanghai Jingyu Intelligent Technology Co.,Ltd,Shanghai 201800,China)

机构地区：[1]武汉大学图像传播与印刷包装研究中心,武汉430079 [2]武汉大学土木建筑工程学院,武汉430079 [3]上海景域智能科技有限公司,上海201800

出　　处：《数字印刷》2022年第2期38-44,60,共8页Digital Printing

摘　　要：颜色恒常性是人类特有的一种心理感知现象,为充分考虑人类在观察外界事物时的心理情绪感知和意识对图像物体或者对象语义的理解,本研究首先基于深度学习模型(Mask-RCNN)对图像内容中的对象类别进行语义分割,构建一个根据对象语义主题划分的图像分类集,在此基础上对经典灰度世界(GW)算法进行改进,将图像对象语义分类集的RGB三通道平均值,作为图像全局光照颜色估计结果,并与GW、WP、SoG和GE四种经典算法的性能进行对比。实验结果表明,本研究算法的还原角度误差的中值误差和均值误差均有明显降低,说明其具有更好的光照颜色估计效果。Color constancy is a kind of psychological perception phenomenon unique to human beings.To fully consider human psychological perception and awareness of the semantics of image objects when observing external things,in this study,firstly object semantic segmentation of image content was carried out based on the deep learning model(Mask-RCNN),and an image classification set was built based on the semantic topics of the objects.And on this basis,the classic Grey World algorithm was improved by using average value of the RGB three-channel the image object semantic classification set as the image global illumination color estimation result.Finally,the performance of the classic algorithms were compared.The results showed that the median error and mean error of the algorithm in this study is significantly reduced,and it has better illuminate estimation effect.

关 键 词：机器视觉 颜色恒常性 光照颜色估计 对象语义 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]