| An image classification system is proposed herein that will recognize objects in images based on specific features of their shapes. The features that the system looks for consist of segments, circles, triangles and quadrilaterals. All objects are assumed to be made up of a combination of the above primitives. Applications of this system are numerous and among them we can mention robot vision, industrial parts manufacturing and medical imaging. The process consists of detecting the edges of the objects, obtaining the curvature of the resulting contour, segmenting the curvature data according to a set of rules and finally extracting the features that will serve to recognize these objects. The feature data consists of 2 elements: a 4 dimensional vector and a string of characters taken from a set of 4 symbols representing each type of primitive mentioned above. A hybrid filter structure is used in edge detection using a combination of linear filters and a median filter. As for the curvature computation a statistical method based on scatter matrix theory is used to overcome the difficulty of defining curvature for discrete curves. The recognition system allows for learning of new objects. |
