Images contain information in a very dense and complex form, which a human eye, after years of training, can extract and understand. The main goal is to extract form an image a set of composing objects or real life attributes. This information is inferred from low-level physical and mathematical properties of the image using a complex model of the reality the image reproduces. The existing systems based on domain specific models are incapable of retrieving images from heterogeneous collections or in the case they are only based on some abstract perceptually semantic feature like texture. Most systems use this lower level approach to retrieve images from heterogeneous collections [4].
Besides containing a large quantity of complex data, images are also of very large dimensionality. Methods like comparing and correlating pixels that operate on the image directly are seldom powerful (costly in terms of time complexity) enough to full the users requirements. The usual approach to overcome this problem is to extract from the image a certain number of relevant features, which reduces the dimensionality, yet preserving useful information. These features are then considered as components of a feature vector, which makes the image, correspond to a point in a much lower but still high dimensional abstract space called the feature space. Two images are then considered similar if their feature vectors lay close in the feature space. The features that are extracted usually fall in three general categories color, shape and texture.