Michaël Antonie van Wyk

THE NORMALIZED WAVELET DESCRIPTOR FOR SHAPE RETRIEVAL

In many image analysis applications, such as image retrieval, the shape of an object is of primary importance. In this paper, a new shape descriptor, namely the Normalized Wavelet Descriptor (NWD), which is a generalization and extension of the Wavelet Descriptor (WD), is introduced. The NWD is compared to the Fourier Descriptor (FD), which in image retrieval experiments conducted by Zhang and Lu, outperformed even the Curvature Scale Space Descriptor (CSSD). Image retrieval experiments have been conducted using a dataset containing 2D-contours of 1400 objects extracted from the standard MPEG7 database. For the chosen dataset, our experimental results show that the NWD outperforms the FD.

Global Image Feature Extraction Using Slope Pattern Spectra

A novel algorithm inspired by the integral image representation to derive an increasing slope segment pattern spectrum (called the Slope Pattern Spectrum for convenience), is proposed. Although many pattern spectra algorithms have their roots in mathematical morphology, this is not the case for the proposed algorithm. Granulometries and their resulting pattern spectra are useful tools for texture or shape analysis in images since they characterize size distributions. Many applications such as texture classification and segmentation have demonstrated the importance of pattern spectra for image analysis. The Slope Pattern Spectra algorithm extracts a global image signature from an image based on increasing slope segments. High Steel Low Alloy (HSLA) steel and satellite images are used to demonstrate that the proposed algorithm is a fast and robust alternative to granulometric methods. The experimental results show that the proposed algorithm is efficient and has a faster execution time than Vincents linear granulometric technique.

Kernel-based non-linear template matching

A new non-linear minimum norm template matching technique is introduced. Similar to the theory of Support Vector Machines the proposed framework is also based on Reproducing Kernel Hilbert Space principles. Promising results when applied to aerial image matching are reported and future work is highlighted.

A GENERALIZED 3-D FOUR-WING CHAOTIC SYSTEM

In this paper, several three-dimensional (3-D) four-wing smooth quadratic autonomous chaotic systems are analyzed. It is shown that these systems have similar features. A simpler and generalized 3-D continuous autonomous system is proposed based on these features which can be extended to existing 3-D four-wing chaotic systems by adding some linear and/or quadratic terms. The new system can generate a four-wing chaotic attractor with simple topological structures. Some basic properties of the new system is analyzed by means of Lyapunov exponents, bifurcation diagrams and Poincare maps. Phase diagrams show that the equilibria are related to the existence of multiple wings.

MODEL-FREE HIGH ORDER DIFFERENTIAL STATES OBSERVER FOR NONLINEAR AFFINE SYSTEM

Abstract This paper considers time-varying nonlinear SISO affine systems with unknown models and unknown bounded disturbance. A high order differentiator is proposed to extract derivatives of the measured signals up to n th -order. Stability and convergence of the differentiator are analyzed. It is also important that the differentiator itself is a stable model-free observer converging to the state evolution of the nonlinear affine system. Simulations verify that it is ideal to estimate the states of nonlinear system and to extract the derivatives of a noisy signal.

An Auction Algorithm for Graph-Based Contextual Correspondence Matching

The Auction Graph Matching (AUGM) algorithm is presented. This algorithm is based on a novel joint probabilistic framework that transforms the graph matching problem into a linear assignment problem which is efficiently solved by the Bertsekas auction algorithm. A salient feature of this single-pass auction-based approach is that the inferred match probabilities are not only constrained over all objects in the reference image, but are also constrained over all objects in the input image.

Generalized Variable-Kernel Similarity Metric Learning

Proximity-based classifiers such as RBF-networks andnearest-neighbour classifiers are notoriously sensitive to the metric used to determine distance between samples. In this paper a method for learning such a metric from training data is presented. This algorithm is a generalization of the so called Variable-Kernel Similarity Metric (VSM) Learning, originally proposed by Lowe and is therefore known as Generalized Variable-Kernel Similarity Metric (GVSM) learning. Experimental results show GVSM to be superior to VSM for extremely noisy or cross-correlated data.