基于共享隐空间的多视角SVM

Multi view SVM based on common hidden space

传统的单视角方法对来自不同场景不同形式的多视角样本难以获得较好的分类性能,因此多视角学习成为近年来的热门研究课题并被广泛研究.在多视角学习中,可能存在这样一种特殊现象,即来自不同视角相同类的样本间的差异比来自同一视角不同类的样本间的差异大,这给多视角学习带来很大挑战,并导致多视角学习效果变差.鉴于此,首先利用Parzen窗技术构建共享隐空间,并将共享隐空间联合原始空间得到扩展空间,进行多视角学习,能够很好应对上述特殊现象;然后利用支持向量机(SVM),提出一种新型的多视角学习方法,即基于共享隐空间的多视角SVM;最后通过在人工和真实的多视角数据集上的实验验证了所提方法在应对上述挑战时具有很好的实验效果.

Because the traditional single-view methods difficultly obtain better classification performance on different scenes and different forms of multi-view samples, multi-view learning has been widely studied and has become one of the hot topics in recent years. However, in multi-view learning, there may be a special phenomenon of that the difference between samples from the same class of different perspectives is larger than that from different classes of the same perspective, which brings great challenges to multi-view learning, and eventually it will lead to poor multi-view learning. The Parzen window technology is used to construct the public space, and the public space is combined with the original space to obtain the extended space for multi-angle learning, so as to meet the challenges brought by the above special phenomena. Then we use the support vector machine(SVM) to propose a kind of new multi-view learning method, namely a multi-view SVM based on shared hidden space. Experiments on real multi-view data sets verify that the proposed method has good experimental results in response to the above challenges.