| 摘要: |
| 不同颜色恒常性算法适用于不同场景下的图像，算法融合是扩展颜色恒常性算法适用范围常用的方法之一，而现有融合性算法在算法选择依据上忽略了语义信息在图像纹理特征描述中的作用，导致光源估计时的精度不高。针对该问题，提出一种语义驱动的颜色恒常决策算法。首先，利用PSPNet(Pyramid Scene Parsing Network)模型对经过一阶灰度边缘算法(1st Gray Edge)偏色预处理后的目标图像进行场景语义分割，并计算场景中各个语义类别的占比；其次，根据语义类别及占比在已训练的决策集合中寻找相似的参考图像，并使用欧氏距离计算两者的语义相似度；最后，将语义相似度与基于多维欧氏空间确定的阈值进行判别，根据判别结果选择合适算法为目标图像实行偏色校正。在Color Checker和NUS-8 camera两种数据集中的实验结果表明，所提算法光源估计角度误差较单一算法均大幅度下降，且较同类型融合性算法分别下降14.02%和8.17%，提高了光源估计的鲁棒性和准确度。 |
| 关键词: 光源估计 图像处理 颜色恒常性 场景语义分割 |
| DOI：10.20079/j.issn.1001-893x.221228001 |
|
| 基金项目:国家自然科学基金资助项目(61307016)；辽宁省创新人才支持计划(LNCX2020005)；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);中国刑事警察学院研究生创新能力提升项目(2022YCYB50) |
|
| A Semantic-driven Color Constancy Decision Algorithm |
| LIU Kai,SUN Peng,TONG Shibo,XIE Mengda |
| (1.Department of Criminal Science and Technique,Criminal Investigation Police University of China,Shenyang 110854,China;2.School of Cyberspace Security,Guangzhou University,Guangzhou 511442,China) |
| Abstract: |
| Different color constancy algorithms are applicable to images of different scenes,and algorithmic fusion is one of the common methods to extend the applicability of color constancy algorithms,while the existing fusion algorithms ignore the role of semantic information in image texture characterization based on algorithm selection,resulting in low accuracy in illuminant estimation.To address this problem,a semantic-driven color constancy decision algorithm is proposed.Firstly,the scene semantic segmentation is performed on the target image after the 1st Gray Edge preprocessing using the Pyramid Scene Parsing Network(PSPNet) model,and the percentage of each semantic category in the scene is calculated.Secondly,similar reference images are searched in the trained decision set based on semantic categories and occupancy ratios and the semantic similarity between them is calculated by using Euclidean distance.Finally,the semantic similarity is discriminated from the threshold value based on multi-dimensional Euclidean space,and a suitable algorithm is selected to implement color bias correction for the target image based on the discriminant result.The experimental results in Color Checker and NUS-8 camera datasets show that the proposed algorithm significantly reduces the angle error of illuminant estimation compared with a single algorithm,and reduces 14.02% and 8.17% respectively compared with the same type of fusion algorithm,which improves the robustness and accuracy of illuminant estimation. |
| Key words: illuminant estimation image processing color constancy scene semantic segmentation |
《电讯技术》编辑部 