A unified approach to feature extraction based on an invertible image transform

Feature extraction is an important component in all areas of image processing a fact demonstrated by the wide variety and diversity of the methods available. These range from statistical to human vision based approaches. Although progress has been fruitful and uninterrupted, it is also apparent that as yet to unified theory of feature identification or representation has emerged. It is towards this goal that this work is directed. The approach adopted has two fundamental principles: meaningful image features are inherently localised in both the spatial and spatial frequency domains; the degree of this locality is not constant across the range of features, in general, image features exist within a multiresolution space. Based on these principles, an attempt is made in this work to provide a unified feature extraction framework. Starting from a general image model, a feature estimation scheme is developed which, by way of example, assumes the image to consist of a multiresolution set of line or edge features. The estimation is achieved by the use of an invertible transform, which by definition incorporates the multiresolution structure underlying the model. The work concludes with a discussion on the appropriateness of the approach to more complex features, such as curvature and shape