多核支持向量域描述在基于图像集合匹配的人脸识别中的应用

Multi-kernel support vector domain description and its application in facial recognition based on image set matching

目的图像集匹配是当前模式识别领域研究的一个热点,其核心问题是如何对图像集合建模并度量两个模型的相似性,为此提出一种基于支持向量域描述的人脸识别的方法。方法支持向量域描述是一种基于支持向量机学习的数据描述方法,可以用于图像集合建模,但是单一的核函数不能准确地描述具有多中心分布的数据。本文通过多核学习扩展了支持向量域描述,提高其对多中心分布数据的表达能力。进一步借助与位置相关的方法对样本动态加权,解决全局权重参数所带来的问题。结果在公开的基于集合的人脸识别数据库上进行测试,在Honda/UCSD、CMU MoBo和YouTube数据库上,本文方法的识别率分别达到100%、98.72%和62.34%。结论实验结果表明,在光照条件受控制的监控环境中,本文方法是有效的,并取得了优于其他基于集合匹配的人脸识别算法。

Objective Image set matching has attracted increasing attention in the field of pattern recognition. For set-based image matching, the key issues can be categorized on the basis of the processes of representing the image set and measuring the similarity between two sets. Method Support vector domain description （SVDD） is a recently developed method based on support vector machine learning. SVDD is a boundary one-class learning method that maximizes the availability of sam- pies that do not belong to the target class in refining its decision boundary, and can be used to describe a set of objects. Accordingly, each image set is described with a hypersphere, and the problem of image set matching is converted into the measure of the clistance between two hyperspheres. Using support vector machine learning, each image set from the original input space is mapped into a high-dimensional feature space and modeled with support vector domain to handle the underly- ing non-linearity in the data space. In the feature space, a hypersphere encloses most of the mapped data. Thereafter, a novel metric is proposed based on domain - domain distance in a high-dimensional feature space; the distance between twoimage sets is then converted into the distance between pair-wise domains. However, the SVDD model has a disadvanta- geously simple form with only a single kernel information. Selecting the best kernel parameters is difficult and the construc- ted hypersphere is considerably sensitive to the trade-off parameter. Multiple kernel learning methods apply multiple kernels instead of merely one specific kernel function and its corresponding parameters. Recent developments in composition kernel learning for classification motivated us to apply a position-based weighting instead of the same global trade-off parameter to discriminate the importance of samples. Furthermore, considering the SVDD model＇ s disadvantageousiy simple form with only one kernel and the difficulty of selecting the best kernel parameters, we propose a muhi-kemel SVDD model, which can flexibly describe the data distribution boundary in the feature space after analyzing the space of multi-kernel mapping. This study utilizes the nearest neighbor classifier to obtain the class label. Result This study＇ s experimental settings reach 100% , 98.72% , and 62. 34% recognition rate in the public HondafUCSD, CMU MoBo, and YouTube video database, respectively. Conclusion Given that multi-kernel learning can improve the efficiency of kernel selection and automatically evaluate the relative importance of the candidate kernels, the multi-kernel SVDD model flexibly describes the data distribu- tion boundary in the feature space and provides a considerably accurate data description for the multifaceted context of the multi-model data set. Experiments conducted on public data sets demonstrate that the multi-kernel SVDD improves predic- tion accuracy and assists in characterizing the properties of complex data.