Efficient image region and shape detection by perceptual contour grouping

Image region detection aims to extract meaningful regions from image. This task may be achieved equivalently by finding the interior or boundaries of regions. The advantage of the second strategy is that once a closure is detected not only its shape information is available, but also the interior property can be estimated with a minimum effort. In this paper, we present a novel method that detects region though region contour grouping based on Generic Edge Token (GET). GETs are a set of perceptually distinguishable edge segment types including linear and non-linear features. In our method, an image is first transformed into GET space on the fly represented by a GET graph. A GET graph presents perceptual organization of GET associations. Two types of perceptual closures, basic contour closure and object contour closure, based upon which all meaningful regions are conducted, are defined and then detected. The detection is achieved by tracking approximate adjacent edges along the GET graph to group the contour closures. Because of the descriptive nature of GET representation, the perceptual structure of detected region shape can be estimated easily based its contour GET types. By using our method, all and only perceptual closures can be extracted quickly. The proposed method is useful for image analysis applications especially real time systems like robot navigation and other vision based automation tasks. Experiments are provided to demonstrate the concept and potential of the method.