Pattern classification and recognition based on morphology and neural networks

Morphological transformations are an efficient method for shape analysis and representation. In this work the pecstrum (pattern spectrum), which is a morphological shape descriptor, is used for object representation. Neural networks are then employed, instead of conventional classification techniques, for object recognition and classification. Various coding schemes and training procedures have been examined in order to achieve a high classification performance. A complete classification and recognition scheme is proposed, which is shown to work satisfactorily even for small objects, where the quantization noise has significantly distorted their shape. The classification results are compared with those obtained using conventional methods, as well as with the results obtained using other shape descriptors.