Feature Based Image Classification in Compressed and Uncompressed Domains

Content-based image retrieval refers to making use of low-level visual features for grouping images into semantically meaningful categories. In this project we aim to study the technique of content-based image retrieval. We shall test the discriminating power of various features to classify images as landscapes and cityscapes by implementing [1]. The features used are: color histograms, color coherence vectors, DCT moments, edge direction histograms and edge direction coherence vectors. In [1] the features are extracted from the raw images (i.e. in the uncompressed domain). We shall attempt to find out how well the classification framework works when instead of using uncompressed raw images, the features are extracted from the compressed-domain images(for e.g. JPEG). Thus all the features mentioned above will be extracted from the block DCT coefficients of the JPEG image. The results obtained in compressed domain and uncompressed-domain are then compared.