Higher-order Co-occurrence Features based on Discriminative Co-clusters for Image Classification

Co-occurrence based image features have attracted keen attentions due to the promising performances for image classification tasks [1, 2, 3, 6, 7]. For extracting the co-occurrences, it is common to transform the quantitative data into qualitative data (symbols) by means of quantization (clustering) at first; e.g., continuous gradient orientation is coded into orientation bins [3], RGB colors are indexed [2] and local features are categorized into visual words [7]. Such point-wise clustering, however, is not necessarily suitable to characterize the pair-wise co-occurrences. And the higher-order co-occurrences beyond pair-wise has been rarely considered due to the exponential increase of the dimensionality by using those point-wise symbols. In this paper, we propose a method to extract image features based on effective higher-order co-occurrences. The proposed method constructs the co-clusters to discriminatively quantize joint primitive quantitative data, such as pair-wise pixel intensities, unlike the standard co-occurrence methods that utilize simple clusters trained in an unsupervised manner for quantizing point-wise data. The discriminative co-clusters effectively exploit the co-occurrence characteristics even by a fewer number of cluster components, resulting in low-dimensional co-occurrence features, which enables us to develop the higher-order cooccurrence features of feasible dimensionality.