The identification of objects in natural images and the semantic interpretation of the image is an important topic of research. For a collection of images, image analysis typically involves a number of steps that employ predetermined tools. The choice of i

recently. Some success with knowledge-based configuration of image processing algorithms [24] has been achieved with systems such as CONNY[29] and SOLUTION[40]. In such systems, image processing operators and their parameters are manipulated till the desired output is achieved. Instead of an exhaustive search by varying all possible combinations of operators and their parameters, a rule base guides the configuration. In several applications dealing with image region identification, the optimisation of low-level operations such as image segmentation, texture characterisation, and classifier choice, are considerably important. Several years of research has produced numerous image segmentation and texture analysis tools and many classifiers to choose from. Our work is based on the premise that for object recognition applications, region-based optimisation of image analysis tools will generate better classification ability with unknown test data. In this paper our main goal is to develop a ‘bank of classifiers’ approach to image region identification and test different classifier (expert) selection schemes. The basic idea is to use an image database that in our case is a collection of natural images since we are primarily interested in natural scene analysis. We train a range of experts on different texture representations of the region, e.g. by using a number of nearest neighbour classifiers, each trained with a separate feature set. We use this ‘bank of classifiers’ in two modes: selective and combinatory. In the selective mode, for a given test region with a possible choice of different texture representations, its statistical properties are used with the knowledge of known class distributions to determine which classifier should be triggered for classification. In this mode a test sample is classified by only one expert. In the combinatory mode, all classifiers are used for classification and results are combined together with well-known classifier combination rules. These two approaches are compared with a baseline approach of a single expert classification for all image regions. This single expert is predetermined either based on experience or prior knowledge of its superiority over other experts.

2.0 Application

This study investigates the application of ‘bank of classifiers’ approach to natural scene analysis application. Even though the methodology has a generic impact on other areas of work related to