Content area
Abstract
This thesis addresses the three major components of a texture classification system: tex-ture image transform, feature extraction/selection, and classification. A unique theoretical investigation of texture analysis, drawing on an extensive survey of existing approaches, defines the interrelations among 11 types of texture analysis methods. A novel unification of the different methods defines a framework of transformation and representation in which three major classes of transform matrices capture texture information of increasing coherence length or correlation distance: the spatial domain method (co-occurrence method), the micro-structural method (run-length method), and the frequency multi-chan-nel method (Fourier spectrum method).
A more concise vector representation of a selected transform matrix is then needed for input to a classifier. Unlike traditional methods, which use various special functions to describe the properties of each transform matrix, a new approach directly applies a princi-ple component analysis technique to the transform matrix. The Karhunen-Loeve Trans-form (KLT) extracts a vector of dominant features, optimally preserving texture information in the matrix. This approach is made possible by the introduction of a novel Multi-level Dominant Eigenvector Estimation (MDEE) algorithm, which reduces the computational complexity of the standard KLT by several orders of magnitude. The statis-tical Bhattacharyya distance measure is then used to rank dominant features according to their discrimination power.
Experimental results of applying the new algorithm to the three transform matrix classes show a strong increase in performance by texture analysis methods traditionally consid-ered to be least efficient. For example, the power spectrum and run-length methods now rank among the best. Using the same MDEE algorithm, the three extracted feature vectors are then combined into a more complete description of texture images. The same approach is also used for a study of object recognition, where the combined vector also include granulometric, object-boundary, and moment-invariant features.
In most classification experiments, a simple statistical Gaussian classifier is used. The plankton object recognition experiments use a Learning Vector Quantization (LVQ) neu-ral-net classifier to achieve superior performance on the highly non-uniform plankton database. By introducing a new parallel LVQ learning scheme, the speed of network train-ing is dramatically increased. Tests show a 95% classification accuracy on six plankton taxa taken from nearly 2,000 images. This result is comparable with what a trained biolo-gist can accomplish by traditional manual techniques, making possible for the first time a fully automated, at-sea approach to real-time mapping of plankton populations.