Higher order spectra invariants for shape pattern recognition

This dissertation addresses the problem of shape feature based invariant pattern recognition. A new integrated feature extraction algorithm is proposed, which includes both the feature generation and feature selection (or reduction) procedures. First, a set of features invariant to rotation, translation, and scaling (RTS) is generated using the Radon transform and bispectral analysis. Such features are called higher order spectra (HOS) invariants in this work. In order to improve the noise resistance of the invariants, the ensemble averaging technique is introduced into the estimation of bispectra. The feature data are further reduced to a smaller set using thresholding and principal component analysis. The resultant feature invariants are proved to be more reliable and discriminable in the classification stage than the initial ones. Further, a class discriminability measure based on the intraclass variance and interclass separation matrix is defined. It is shown that the large values of this measure always indicate the high classification power of the derived features.; A pattern recognition system is then developed based on the proposed feature extraction algorithm. It is shown that simple training and classification processes are applicable since the extracted HOS invariants form compact and isolated clusters in the feature space. The experimental results show that a variation version of minimum distance classifier yields high classification accuracy even with low SNR inputs. The comparison study with other two well-established methods, Hu's moment invariants and Fourier descriptors also shows that the performance of the proposed method is much favorable to those two methods especially in the presence of background noise.; Finally, the application of the proposed algorithm in multimedia and imaging database systems is addressed. A simplified version of HOS invariants is used for image indexing in the developed image storage and retrieval system. A new similarity matching technique based on Tanimoto measure is introduced for fast image retrieval. The system retrieval accuracy is high in the presence of noisy query images with low SNR's as shown in the simulation results.