Compressed image classification using bag of visual words

In multimedia applications image and video retrieval techniques are crucial and for efficient storage and transmission, most multimedia information is in compressed form. The main issue is to extract features which are invariant to rotation, scale changes, view-point change, and local affine transformations from the compressed image which is the fundamental step towards content-based retrieval of image data. In this paper a model based on Bag of Visual words for image classification is implemented in which features like SIFT and SURF are extracted from images which are compressed and decompressed using various well known compression schemes namely DCT Discrete Cosine Transform, AMBTC Absolute Moment Block Truncation Coding and Sub-band thresholding with Hufftnan Encoding. The decompressed set of images is used as the input to object recognition process. In this model the key points in the image are mapped into visual words and the image is represented as Bag. of Visual words used as feature vectors in the classification task. A performance evaluation on compressed images of Caltech database under SURF and SIFT descriptors with codebook created using K-means clustering and classification using KNN classifier has been done by varying the dictionary size.