基于距离约束稀疏/组稀疏编码的自动图像标注

Distance Constraint Sparse/Group Sparse Coding for Automatic Image Labeling

为解决图像自动标注中的语义鸿沟问题,有效选择并利用图像特征,提出基于距离约束稀疏/组稀疏编码(distance constraint sparse/group sparse coding,DCSC/DCGSC)的2种特征选择算法,并分别应用到图像自动标注任务中。考虑到不同特征基相似性对图像语义相似性的贡献不同,定义了度量二者相关性的距离约束正则项。将该正则项分别集成到稀疏/组稀疏编码的特征选择模型中,使选择的特征在保证稀疏性/组稀疏性的同时,优先选择与语义相似性描述最接近的视觉特征基。利用在训练图像集中学习的特征权值,寻找测试图像的K最近邻(Knearest neighbor,KNN)图像,并通过标签转移实现图像标注。在Corel5K图像库上测试标注性能,集成多特征的DCGSC查准率、查全率和标注正确的关键词个数可达32%、34%和151,优于其他相关标注算法。而对于单特征图像,使用DCSC也能改善标注性能。可见,距离约束对特征选择和图像标注是有效的。

In order to bridge the semantic gap in automatic image labeling, and effectively leverage image features, two feature selection algorithms based on distance constraint sparse/group sparse coding （DCSC/DCGSC） were presented to solve the problem of image se- mantic labeling. Considering that feature atoms similarity may have different contribution to the serhantie similarity between images, a distance constraint regularization was defined and integrated with sparse/group sparse coding for feature selection, which encourages the feature atoms with sparsity/group sparsity and more Similar to the semantic discrimination to be enforced. Given a test image, the K-Nea- rest Neighbors （KNN） can be found using the learned feature weights from the training images and labels can be transfered. Experimental results on Corel5K showed that DCGSC outperforms other related method with the average precision of 32% , average recall of 34% , and the numbers of total labels recalled of 151. For images represented with single type of feature, DCSC also helps to improve the annotation performance, which validates the effectivity of distance constraint for image labeling.