Content representation and similarity matching for texture-based image retrieval

This paper addresses the fundamental issues of visual content representation and similarity matching in content-based image retrieval and image databases in general. Simply stated, defining an image retrieval system is equivalent to find answers to two fundamental questions: 1. Representation model or which features are used to represent the content of images; 2. Once the set of features representing the content of images is determined, the question of how to combine the individual or partial similarities according to each feature to form a global similarity must be addressed. In this paper, a new similarity model is introduced based on the Gower coefficient of similarity. This similarity model is flexible and can be declined in several versions: non-weighted, weighted and hierarchical versions. This model was applied to a sample of homogeneous textured images considering two representation models: the autoregressive model, a purely statistical model, and an empirical perceptual model based on perceptual features such as coarseness and directionality. Experimentations show very interesting results.