Spatial Graphical Models for Image Matching and Object Recognition

摘要

In this project we have had two partial successes. First is an efficient detection algorithm for objects in complex scenes, using very simple spatial arrangements to represent the objects, based on local features which are automatically identified in training. The simplicity of the arrangement allows us to use the Hough transform to very quickly find a small number of candidate locations for the objects. We have also proposed a parallel architecture for implementing this algorithm with interesting biological analogies. Second is an algorithm for isolated object recognition using decision trees to gradually explore the natural partial ordering of the space of spatial arrangements. The principles of this algorithm have also been successfully applied to the recognition of acoustic signals.