Image retrieval using compression-based techniques

A new method for semantic retrieval of color images employing data compression is presented. While typical content-based image retrieval systems operate in some parameter space, the introduced data-driven technique uses as features the very image data, represented as sets of recurring patterns collected in dictionaries. In a first offline step, the images are quantized in the Hue Saturation Value space and converted into strings, after being modified to preserve some vertical information in the process, and representative dictionaries are extracted from each string with a data compression algorithm; subsequently, the dictionaries are matched in pairs and the distance between each couple of them is estimated. On the basis of the computed distances, the system enables the user to retrieve images with similar content to a given query image. Compression-based classification techniques, being data-driven, have the drawback of being computationally intensive and have been applied, in the general case, to restricted datasets; instead, the low-complexity solutions employed in this work allow applying this similarity measure on larger datasets, keeping at the same time the desirable parameter-free approach which is characteristic of these methods. Experiments show that the proposed technique outperforms similar recent concepts based on Vector Quantization.