Recognition of Similar Objects Using a Hybrid Classifier

A novel method for classification of objects based on a hybrid of decision theoretic and structural methods is presented in this paper. The circumstances of scaling and presence of noise are included as the part of the study. Images are degraded by some known types of noise like Gaussian, salt & pepper and speckle and iterative filtering algorithm based on classification results and using alpha-trimmed mean filter was used. Fuzzy clustering algorithm is used for thresholding and background removal in cluttered images and spurious parts are reduced using morphological operations. Input database is made up of images having similar shapes lied on their most usual appearance. Feature vectors are composed of moment invariant and Interior angles of polygons and was extracted after normalizing the object boundary with respect to size and orientation. Interior angles are extracted from a shape, described using a new polygonal approximation technique. Similarity measurement is done by combining two classifiers, Euclidean distances in decision theoretic and string matching in structural methods. In order to investigate the reliability of presented method in presence of noise, the classification results obtained from a hybrid method are compared with those of the decision theoretic or structural methods.