An association architecture for the detection of objects with changing topologies

This paper presents an architecture for image analysis that is based on feature hierarchies. The architecture allows for shift, scale and topological invariant detection of objects. Features are efficiently represented and combined dynamically during the detection process. The respective feature detectors are trained using a supervised learning scheme. The method discussed here can also solve the problem of segmenting an image into image regions that correspond to detected features. This segmentation can be done through backtracking of feature information in the feature hierarchy. We applied the method for a set of images where building facades are analyzed and show experimental results that demonstrate the capabilities of the system.