Visual Recognition Using Density Adaptive Clustering

Visual codebook based texture analysis and image recognition is popular for its robustness to affine transformation and illumination variation. It is based on the affine invariable descriptors of local patches extracted by region detector, and then represents the image by histogram of the codebook constructed by the feature vector quantization. The most commonly used vector quantization method is k-means. But due to the limitations of predefined number of clusters and local minimum update rule, we show that k-means would fail to code the most discriminable descriptors. Another defect of k-means is that the computational complexity is extremely high. In this paper, we proposed a nonparametric vector quantization method based on mean shift, and use locality-sensitive hashing (LSH) to reduce the cost of the nearest neighborhood query in the mean-shift iterations. The performance of proposed method is demonstrated in several image classification tasks. We also show that the Information Gain or Mutual Information based feature selection based on our codebook further improves the performance.