Mykhaylo Yatsymirskyy

Fast orthogonal neural networks

The paper presents a novel approach to the construction and learning of linear neural networks based on fast orthogonal transforms. The orthogonality of basic operations associated with the algorithm of a given transform is used in order to substantially reduce the number of adapted weights of the network. Two new types of neurons corresponding to orthogonal basic operations are introduced and formulas for architecture-independent error backpropagation and weights adaptation are presented.

Efficient 1D and 2D Daubechies Wavelet Transforms with Application to Signal Processing

In this paper we have introduced new, efficient algorithms for computing one- and two-dimensional Daubechies wavelet transforms of any order, with application to signal processing. These algorithms has been constructed by transforming Daubechies wavelet filters into weighted sum of trivial filters. The theoretical computational complexity of the algorithms has been evaluated and compared to pyramidal and ladder ones. In order to prove the correctness of the theoretical estimation of computational complexity of the algorithms, sample implementations has been supplied. We have proved that the algorithms introduced here are the most robust of all class of Daubechies transforms in terms of computational complexity, especially in two dimensional case.

Frequency Domain Methods for Content-Based Image Retrieval in Multimedia Databases

Content-based image retrieval is an important application area for image processing methods associated with computer vision, pattern recognition, machine learning and other fields of artificial intelligence. Image content analysis enables us to use more natural, human-level concepts for querying large collections of images typically found in multimedia databases. Out of the numerous features proposed for image content description those based on frequency representation are of special interest as they often offer high levels of invariance to distortions and noise. In this chapter several frequency domain methods designed to describe different aspects of an image, i.e. contour, texture and shape are discussed. Current standards and database solutions supporting content-based image retrieval, including SQL Multimedia and Application Packages, Oracle 9i/10g interMedia and MPEG-7, are also presented.

Fast Orthogonal Neural Network for Adaptive Fourier Amplitude Spectrum Computation in Classification Problems

Fourier amplitude spectrum is often applied in pattern recognition problems due to its shift invariance property. The phase information, which is frequently rejected, may however be also important from the classification point of view. In this paper, fast orthogonal neural network (FONN) is used to compute amplitude spectrum in an adaptable way, enabling to extract more class-relevant information from input data. The complete architecture of the neural classifier system is presented. The proposed solution is compared to standard multilayer perceptron on an artificially generated dataset, proving its superiority in terms of computational complexity and generalization properties.

On Feature Extraction Capabilities of Fast Orthogonal Neural Networks

The paper investigates capabilities of fast orthogonal neural networks in a feature extraction task for classification problems. Neural networks with an architecture based on the fast cosine transform, type II and IV are built and applied for extraction of features used as a classification base for a multilayer perceptron. The results of the tests show that adaptation of the neural network allows to obtain a better transform in the feature extraction sense as compared to the fast cosine transform. The neural implementation of both the feature extractor and the classifier enables integration and joint learning of both blocks.

Fast Adaptive Fourier Transform for Fourier Descriptor Based Contour Classification

In this paper the authors presented the results of experiments on evaluation of efficiency of fast adaptive Fourier transform algorithm applied to the tasks of Fourier descriptor based contour classification. The experiments involved popular SQUID boundaries database of sea fish of different species. For object classification a nonparametrised nearest-neighbor classifier was used and distances between feature vectors were expressed in Euclidean metric.

Neural Network in Fast Adaptive Fourier Descriptor Based Leaves Classification

In this paper the results in leaves classification with non-parametrized one nearest neighbor and multilayer perceptron classifiers are presented. The feature vectors are composed of Fourier descriptors that are calculated for leaves contours with fast adaptive Fourier transform algorithm. An application of fast adaptive algorithm results in new fast adaptive Fourier descriptors. Experimental results prove that the fast adaptive Fourier transform algorithm significantly accelerates the process of descriptors calculation and enables almost eightfold reduction in the number of contour data with no effect on classification performance. Moreover the neural network classifier gives higher accuracies of classification in comparison to the minimum distance one nearest neighbor classifier.

Application of Multiple Sound Representations in Multipitch Estimation Using Shift-Invariant Probabilistic Latent Component Analysis

Probabilistic analysis has become one of the most important directions for development of new methods in Music Information Retrieval MIR field. Its ability to correctly find necessary information in the music audio recordings is especially useful in multipitch estimation, a vital task belonging to the MIR field. Since the multipitch estimation is still far from being resolved, it is important to enhance the existing state-of-the-art methods. Usually, a spectrogram, generated from the Constant-Q transform CQT is used as a basis for the SI-PLCA method. The new approach involves application of more than one method cepstrum and CQT in association of the shift-invariant probabilistic latent component analysis approach and additional processing of all the sound representations, in order to achieve better results.