Gaussian mixture model with semantic distance for image classification

This paper mainly introduces a new method for image classification. The traditional Bag of Visual Words model (BoVW) is a promising image representation technique for image classification. But its limitation is that much valuable information is lost when building the codebook of BoVW simply by clustering visual features in the Euclidian space. In this paper, we take full advantage of image semantic information to learn a new distance metric which achieving the minimal loss of image information, and then we learn visual words by clustering the local features using Gaussian Mixture Models (GMMs) with this distance metric. When given a test image, it firstly forms a visual document using GMM based on our learned distance metric, then its category is determined by estimating the maximum probability using language model under a specific category. Experimental results confirm the effectiveness of our method, results are satisfactory and competitive compared with traditional and state of the art methods.