Contextual and Non-Combinatorial Feature Extraction: Augmented Feature Representation and Its Application to Shape Representation

Extracting features from an image is the first step in many computer vision applications. Traditionally, features represent physical or visual primitives such as edges and corners. In this paper, we augment the definition to include any attributes that conveniently describe the correlation and contextual relation between the primitive and its neighbors. The augmentation allows us to design a more detailed probability distribution model. If the distribution model is differentiable with respect to each attribute of a feature, a simple local search will find a feature set that is a local maximum in the joint probability distribution. Therefore, the final representation is free from noise and aliasing that perturbs the representation away from the local maximum. We can apply the approach to many low level vision tasks. In this paper, we demonstrate our approach with sub-pixel contour representation and surface representation problems.