Over the past decade, research into computer vision has proliferated with the goal to incorporate artificial intelligence into a wide range of applications. These applications can be as sophisticated as intelligent assistants in smartphones and self-driving cars or as mundane as text and face recognition. While most of these applications are software based, they represent unique challenges when it comes to industrial implementation. This thesis concentrates on an optical non-destructive testing (NDT) and automatic classification methodology using customized image processing techniques. In contrast to conventional spatial analyses, which are highly susceptible to noise and human perception, our proposed transform domain approach provides a high degree of robustness and flexibility in feature selection and hence a better classification efficiency. Our presented algorithm classifies the Part-Under-Test (PUT) into two bins of either acceptable or faulty using transform domain techniques in conjunction with a classifier. Because the classification is critically dependent on the features extracted from these images, a sophisticated scalable database was created. This thesis applies transform domain techniques such as Discrete Wavelet Transform (DWT) and Rotated Wavelet Transform (RWT) for feature extraction and then classifies the PUT based on those features. Although, this approach achieves promising classification efficiency, it does not meet industrial standards. It was concluded that in order to achieve those standards, the effect of emissivity fluctuations of the PUT should be negated. The research was then extended to apply an image acquisition algorithm in the form of shape from polarization. The approach exploits the partially linearly polarization of reflected light from the PUT surface. It was observed that this method could not only detect if the PUT is faulty or fault free, but also highlight the locations of the flaws.

Recently, pattern recognition/classification has received considerable attention in diverse engineering fields such as biomedical imaging, speaker identification, fingerprint recognition, and face recognition, etc. This study contributes novel techniques for facial image recognition based on the Two dimensional principal component analysis in the transform domain. These algorithms reduce the storage requirements by an order of magnitude and the computational complexity by a factor of 2 while maintaining the excellent recognition accuracy of the recently reported methods. The proposed recognition systems employ different structures, multicriteria and multitransform. In addition, principal component analysis in the transform domain in conjunction with vector quantization is developed which result in further improvement in the recognition accuracy and dimensionality reduction. Experimental results confirm the excellent properties of the proposed algorithms.

No description available.

No description available.

Transform-domain (TD) sampling is seen as a potential candidate for wideband and ultra-wideband high-performance receivers and is investigated in detail in this research. TD receivers expand the signal over a set of basis functions and operate on the digitized basis coefficients. This parallel digital signal processing relaxes the sampling requirements opening the doors to higher dynamic range and wider bandwidth in receivers. This research is focused on the implementation of a high performance multi-channel wideband receiver that is based on Frequency-domain (FD) sampling, a special case of TD sampling. To achieve high dynamic ranges in these receivers, it is critical that the digital post processing block matches the analog RF front end accurately. This accurate matching has to be ensured across several process variations, mismatches and o�sets that can be present in integrated circuit implementations. A unified model has been defined for the FD multi-channel receiver that contains all these imperfections and a joint synchronization and calibration technique, based on the Least-mean-squared (LMS) algorithm, is presented to track them. A maximum likelihood (ML) algorithm is used to estimate the frequency offset in carriers which is corrected prior to LMS calibration. Simulation results are provided to support these concepts. The sampling circuits in FD receivers are based on charge-sampling and a multi-channel charge-sampling receiver creates an inherent sinc filter-bank that has several advantages compared to the conventional analog filter banks used in other multi-channel receivers. It is shown that the sinc filter banks, besides reduced analog complexity, have very low computational complexity in data estimation which greatly reduces the digital power consumption of these filters. The digital complexity of data estimation in the sinc fiter bank is shown to be less than 1=10th of the complexity in analog filter banks.

The problem of determining an objective means for assessing local characteristics of television subpictures in a Hadamard transform multicoder scheme is stated and discussed. Detailed investigations of transform domain coefficient statistical characteristics for different test images have been conducted. Both monochrome and colour signals were used, as well as different transform sizes and shapes. " Directing Indexes " are proposed which, depending on the inter-relationships among transform coefficients and groups of coefficients, direct each subpicture to the appropriate coder. Three indexes in the case of monochrome signals are proposed, each with its own computational procedures and application requirements. Necessary modifications and changes for application of some indexes on colour signals are also discussed. The proposed technique of indexing eliminates the necessity of equal distribution of subpictures among 'activity classes', a major disadvantage encountered in present activity index. Coders to be used with each directing index are devised and tested, subject to an arbitrary bit rate of 2 bits per pixel, with satisfactory performance compared with some published results for other techniques.

No description available.

The problem of determining an objective means for assessing local characteristics of television subpictures in a Hadamard transform multicoder scheme is stated and discussed. Detailed investigations of transform domain coefficient statistical characteristics for different test images have been conducted. Both monochrome and colour signals were used, as well as different transform sizes and shapes. " Directing Indexes " are proposed which, depending on the inter-relationships among transform coefficients and groups of coefficients, direct each subpicture to the appropriate coder. Three indexes in the case of monochrome signals are proposed, each with its own computational procedures and application requirements. Necessary modifications and changes for application of some indexes on colour signals are also discussed. The proposed technique of indexing eliminates the necessity of equal distribution of subpictures among 'activity classes', a major disadvantage encountered in present activity index. Coders to be used with each directing index are devised and tested, subject to an arbitrary bit rate of 2 bits per pixel, with satisfactory performance compared with some published results for other techniques.

No description available.

No description available.