Design of optimal subband filter banks for image discrimination

The primary objective of this work is to improve texture classification system performance. The work is extended to improve the accuracy with which the faulty components of a printed circuit board are detected from a video sequence of infrared images generated by warming the board at power up. The direct motivation of this research is to enhance the FAULT DETECTION and IDENTIFICATION (FDI) system performance based on classification of the components in the circuit boards. The classification problem may be divided into the stages of feature extraction, dimensionality reduction and pattern recognition. Central to this work is that the signal representation plays a crucial role in the classification performance. Specifically, it is proposed that designing an optimal sub-band filterbank for fault detection and identification or texture classification improves the classification performance when the filterbank is used for that purpose.; The focus of this dissertation is on the design of subband filterbanks for feature extraction and classification of images. One of the major conclusions of the experiments is that the wavelet used for decomposing the images for classification plays a crucial role in the classification task. Furthermore, the commonly used octave band decomposition is evaluated against alternative decompositions. It is concluded that non-octave decompositions are generally superior. Also, the classification performance using various feature extraction techniques along with dimensionality reduction methods are compared. A quadrature mirror filterbank designed is tested in texture classification and fault detection, and results in superior classification performance compared to other filterbanks.; Optimal filters designed with image compression in mind do not guarantee optimality with respect to discrimination. Therefore, approaches for the design of optimal filterbanks with optimal discrimination are proposed. A simulated annealing algorithm is used to find the optimal filter coefficients by maximizing class separability. Algorithms are developed to find the optimal filterbank for a given dataset and to classify an unknown texture or to find if the given component is faulty or not.; Performance of the proposed methods is demonstrated in extensive experiments, which justify the new approaches.