A Coupled Relaxation Method for Finding Perceptual Structures
In this paper we describe a method for determining the existence and parameters of a geometric structure in an image by using a relational representation of its structure. The relational model is matched to image structure in order to find possible instances of the model in the image. Matches between the relational model and image primitives are then used to determine probability distributions for the parameters of a geometric transformation. This transform maps a geometric model of the structure onto an instance in the image. This distribution may then be used to infer a probability map for pixel-based information such as edge responses. When combined with the original edge responses, an enhanced image is produce with more salient edge structure. Iteration of the procedure results in a consistent set of models and edge structure. The method is demonstrated on rectangles undergoing affine transforms.