Fast multi-view segment graph kernel for object classification

Object classification is an important issue in multimedia information retrieval. Usually, we can use images from multiple views (or multi-view images) to describe an object for classification. However, two issues remain unsolved. First, exploiting the spatial relations of local features from different view images for object classification. Second, accelerating the multi-view object classification process. To solve these two problems, we propose fast multi-view segment graph kernel (FMSGK). Given a set of multi-view images for an object, we segment each of them in terms of its color intensity distribution. And inter- and intra-view segment graphs are constructed to describe the spatial relations of the segments between and within view images respectively. Then, these two types of graphs are integrated into a so-called multi-view segment graph. And the kernel between objects is computed by accumulating all matchings' of walk structures between their corresponding multi-view segment graphs. Since computing the kernel directly is highly time-consuming, an accelerating algorithm is derived. Finally, a multi-class support vector machine (SVM) (Duda et al., 2000 [19]; Wang et al., 2008 [32]; Dai and Mai, 2012 [6]) is trained based on the computed kernels for object classification. The experimental results on three data sets validate the effectiveness of our approach.

► Multi-view segment graph kernel is proposed for multi-view images representation. ► Two graphs to describe spatial relations of segments between and within multi-view images. ► An algorithm to accelerate the multi-view segment graph kernel computation.