Multi-local model image set matching based on domain description

• We propose a novel multi-local mode image set matching algorithm.
• It is the first attempt to introduce the SVDD to image set matching.
• We solve the set matching as an optimization problem via domain description.
• We formulate the image set matching as the distances between pair-wise domains.
• Experimental results demonstrate the superior performance of our approach.

Image set matching attracted increasing attention in the field of pattern recognition. Recently, there are a number of effective image set-based matching methods under controlled environment. However in the more complex environment, like multi-view and illumination changed, it is still a challenging problem to develop unsupervised image set matching method to handle multi-local model data. To solve this problem, in this paper, we present a novel multi-local model image set matching method based on data description techniques. First, every image set is divided into multi-local models, and each local model corresponds to a data domain, that is, we innovatively train a support vector data domain to describe each local model by means of the excellent data description ability of support vector data domain, hence each image set can be expressed by a plurality of support vector data domain. Second, a new similarity measure based on domain–domain distance is proposed, and then the distance between two image sets is converted to integrate the distance between pair-wise domains. Finally, the proposed method is evaluated on both set-based face recognition and object classification tasks. Extensive experimental results show that the proposed method outperforms other state of the art set-based matching methods in three public video databases.