Image model based on salient regions and its applications

Detection, representation, and training are the three main issues that need to be resolved in an object recognition or classification system. One possible method is using collection of regions to represent object categories, where each region has a distinctive feature. In this paper we present a region-based image model which will learn and classify objects by training the image model with variant of the objects within the same category. Each object category is represented by a constellation of representative parts. These regions are detected by salient region detector over suitable scales. The standard disjunction rules is applied to construct the image model. During the learning procedure, the distance between any two regions is calculated and accumulated as a measure which is inversely proportional to the probability of a match. The regions with large distances will be removed from the image model iteratively. Finally a small set of regions will be kept as the image model. This image model can be used to retrieve similar images or for object classification. Experimental results show the method is easy to calculate and efficient.