Over-segmentation of edge features has been a challenging problem for many edge-based vision applications. Too many useless features are simply background noise which are costly for higher-level processing. The conventional methods of dealing with oversegmentation use various noise suppressing filters at pixel level for the entire image, and then form features by grouping identified edge points. The computation cost and lack of global heuristics are the major drawbacks. This paper presents a perceptual organization based method for both noise suppression and perceptual feature classification of pre-tracked and partitioned edge segments. In this method, edge traces are selectively tracked and partitioned based on a previously proposed model. Each partitioned edge segment is then classified into noise or perceptual features using the continuity of gradient distribution along the segment. Noise segments can be identified by their discontinuous gradient magnitude. The remaining segments can be further classified into curve and straight line features by identifying constant or monotonic changes of gradient direction. Because its only computations are simple statistics based on a small edge data set, our method is well suited for realtime or hardware-limited applications. Experiments and analysis are provided.
Keywords:

Edge feature, Feature extraction, Over-segmentation, Noise removal, Noise depression, Feature classification, Perceptual organization.