Skip to Main Content
In this paper, we address the issue of feature compression for distributed image classification systems. Such systems often extract a set of features, such as color, texture and shape, from the raw image data and store them as content descriptors. Users on the network search and exchange information with each other using feature data that contains sufficient information for image classification/retrieval. A query is formed at the client by extracting the features from the sample image, having them compressed and transmitted to the remote server. The corresponding image is classified through comparing the query feature vector with all the entries in the database and finding the most similar ones. We pose this problem as a statistical classification problem and consider the design of a transform coding scheme minimizing the detrimental effect of compression. Different from traditional transform coding which was designed to provide the best reconstruction from compressed data, here the goal is to preserve the separation between the feature vectors generated by different classes. Linear discriminant analysis is used due to the property that discrimination information is preserved in lowest dimensional space in transform domain, so that efficient scalar quantization and entropy coding can be applied to the transform coefficients. We propose a greedy algorithm to perform bit allocation in transform domain such that classification information from all dimensions can be employed in quantizer design. We show that our proposed transform coding scheme achieves much better classification performance than traditional KLT transform coding, based on both real data (Brodatz textures) and synthesized random source.