场景语义引导下的改进灰度世界算法  

Improving Gray World Algorithm Guided by Scene Semantics

作　　者：解梦达 孙鹏[1,2] 郎宇博 魏千杰 单大国 Xie Mengda;Sun Peng;Lang Yubo;Wei Qianjie;Shan Daguo(Department of Criminal Science and Technique,Criminal Investigation Police University of China,Shenyang 110854;Key Laboratory of Forensic Science,Ministry of Justice,Shanghai 200063)

机构地区：[1]中国刑事警察学院公安信息技术与情报学院,沈阳110854 [2]司法部司法鉴定重点实验室,上海200063

出　　处：《计算机辅助设计与图形学学报》2022年第1期152-162,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划专项(2017YFC0822204);公安部技术研究计划(2020JSYJC25);司法部司法鉴定重点实验室开放课题(KF202014);辽宁省创新人才支持计划(LNCX2020005);辽宁省教育厅青年科技人才“育苗”项目(JYT2020130);中国刑事警察学院研究生创新能力提升项目(2020YCZD03).

摘　　要：色彩恒常性的实现中,灰度世界算法通常因自然图像间场景差异导致较弱的光源估计性能.针对上述问题,提出了一种以场景语义为引导的改进灰度世界算法.首先,使用灰度化图像计算稠密SIFT描述子以避免彩色图像偏色干扰,SIFT描述子集合由词袋模型进一步生成无序视觉词汇;其次,以视觉词汇对应像素平均亮度为权重,基于空间金字塔匹配算法构建金字塔结构的亮度加权词频直方图;然后,采用直方图交叉核函数计算图像间场景语义相似度,并依据该相似度检索与测试图像场景语义相似的候选图像集;最后,所有候选图像的像素统计分布在经孤立森林算法移除异常值后被用于自适应推断,并更新灰度世界算法的固定假设,改进后灰度世界算法通过光源估计实现色彩恒常性.在ColorChecker,Cube+和NUS共3类公开色彩恒常数据集中的实验结果表明,所提算法在单相机测试中的光源估计角度误差较同类型灰度世界改进算法下降接近20%,且可在跨相机测试中取得对比算法中最低的光源估计角度误差.Gray world algorithm typically shows poor illuminant estimation performance due to scene variations of natural images.To address the above problem,an improved gray world algorithm guided by scene semantics is proposed.Firstly,dense-SIFT descriptors are calculated using grayscale images to avoid the interference of color-biased images,followed by the bag of words(BoW)model to generate unordered visual vocabulary.Secondly,based on the average brightness of these visual words,word frequency histograms are constructed weighted by brightness using the spatial pyramid matching(SPM)algorithm.Thirdly,the scene semantics similarity between images employing the histogram cross kernel function,and retrieve a set of candidate images that similar to the test image is calculated.Finally,after removing the outlier images using the isolated forest algorithm,the pixel statistics distributions of all remaining images are used to adaptively infer and update the fixed assumption of the gray world algorithm.The improved gray world algorithm implements color constancy through illuminant estimation.Experimental results in three publicly available color constancy datasets(ColorChecker,Cube+and NUS)show that the proposed algorithm outperforms comparable improved gray world algorithms in single camera tests(nearly 20%angular error improvement),at the same time,achieves optimal performance in cross-camera tests.

关 键 词：色彩恒常性 灰度世界算法 光源估计 场景语义 SIFT描述子 空间金字塔匹配 孤立森林 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]