Multi-layer Orthogonal Visual Codebook for Image Classification

Recently, Bag of Visual Words (BoW) model has shown its success in image classification and retrieval. The key idea behind the BoW model is to quantize the continuous highdimensional space of image features (e.g. SIFT [1]) to a manageable visual codebook. The quality of the visual codebook has an important impact on BoW-based methods. Different from the existing techniques, such as Kernel codebook [4] and Sparse Coding [5], we propose a novel multi-layer orthogonal codebook (MOC) generation approach. It aims at explicitly reducing quantization error by generating codebook from the feature space that is orthogonal to the existing codebook. Furthermore, we propose two simple schemes to apply the new codebook with spatial pyramid matching (SPM) [2] on image classification. Experimental results show the efficiency of our proposed MOC generation method, and the performance on image classification is comparable to the state-of-the-art techniques.