This doctorate thesis constitutes an approach to improve and extend the state-of-the-art methods for automatic content-based object recognition and retrieval. A detection and segmentation method is proposed for obtaining the image areas of interest, based on image symmetry and phase congruency calculations. Also a method for color recognition, based on histograms, is presented. Invariant features (keypoints) are detected and extracted using a scheme based on scale-space theory. The invariant features are matched with ones from a database of model features.
The method of Merged Feature Matching is proposed and it is based on keypoint fusion from multiple views for the same object scene. This method is developed and tested successfully with increased recognition robustness. A SIFT-based method (V-SIFT) without rotation invariance is employed, in order to speed-up the keypoint extraction and description process.
Methods for feature searching and matching are tested and a data structure based on a KD-Tree is employed, as the most efficient. Matching results are subsequently tested for geometrical consistency, using a generalised Hough transform (GHT) for feature clustering and RANSAC method for query and model images homography estimation. In addition, the proposed methods of K-neighbours GHT, the characteristic reduced total distance and the neighbours fleet geometry, are developed with encouraging results.
An experimental validation of the method is given, with successful vehicle recognition. The proposed method is compared with a probabilistic artificial neural network approach, with higher performance in terms of precision and efficiency. Objects are recognised even with partial occlusion from other objects in the image scene. The proposed method could also be applied in automatic medical and satellite image recognition, image panorama stitching, automatic robot navigation etc.
Key Words : Invariant, features, characteristics, image, recognition, scale-space, searching, matching, vehicle.