Sudhakar, Raghavan

A new approach to estimating convolved signals, refered to as homomorphic estimation, is presented. This method is the fusion of two well-developed signal processing techniques. The first is the class of homomorphic systems which are characterized by a generalized principle of superposition and allow any linear filter method to be applied when signals are non-additively combined. The second well-known technique is the Wiener estimation filter which has the ability to estimate a desired signal in the presence of additive noise. The theory and realization of the homomorphic system for convolution based on the Fourier transform is developed. Homomorphic estimation system performance is analyzed using digital computer simulation. Homomorphic detection is also presented and is shown to be a useful and easily implemented method.

This thesis deals with a study of using the stereo vision technique in the robot calibration. Three cameras are used in measurement to extract the position information of a target point attached onto each of the robot manipulator links for the purpose of identifying the actual kinematic parameters of every link of the robot manipulator under testing. The robot kinematic model used in this study is the S-Model which is an extension of the well-known Denavit-Hartenberg model. The calibration has been done on the wrist of the IBM 7565 robot. The experiment set-up and results and the necessary software are all presented in this thesis.

An IBM Personal Computer AT system was used in conjunction
with a professional data acquisition unit to digitize and
store several cassette deck recorded piano notes. These
digitized notes were then visually analyzed both with an AT
monitor and a high resolution plotter. Fourier and Walsh
Transformations were then performed on the digitized data
to yield further visual information. Upon completion of
this visual study, several types of data reduction and
waveform synthesis methods were formulated. These
experimental methods tested included a wide range of signal
processing techniques such as Fourier Transformation, Walsh
Transformation, Polynomial Curve Fitting, linear Interpolation,
Amplitude Normalization, and Frequency Normalization. The
actual test performed on the experimental synthesis method
consisted of recreating the piano note and then subjectively
comparing the audio performance of the synthetic note versus
the original note.

The main objective of the research is to develop computationally efficient hybrid coding schemes for the low bit implementations of image frames and image sequences. The basic fractal block coding can compress a relatively low resolution image efficiently without blocky artifacts, but it does not converge well at the high frequency edges. This research proposes a hybrid multi-resolution scheme which combines the advantages of fractal and DCT coding schemes. The fractal coding is applied to get a lower resolution, quarter size output image and DCT is then used to encode the error residual between original full bandwidth image signal and the fractal decoded image signal. At the decoder side, the full resolution, full size reproduced image is generated by adding decoded error image to the decoded fractal image. Also, the lower resolution, quarter size output image is automatically given by the iteration function scheme without having to spend extra effort. Other advantages of the scheme are that the high resolution layer is generated by error image which covers the bandwidth loss of the lower resolution layer as well as the coding error of the lower resolution layer, and that it does not need a sophisticated classification procedure. A series of computer simulation experiments are conducted and their results are presented to illustrate the merit of the scheme. The hybrid fractal coding method is then extended to process motion sequences as well. A new scheme is proposed for motion vector detection and motion compensation, by judiciously combining the techniques of fractal compression and block matching. The advantage of this scheme is that it improves the performance of the motion compensation, while keeping the overall computational complexity low for each frame. The simulation results on realistic video conference image sequences support the superiority of the proposed method in terms of reproduced picture quality and compression ratio.

The dual issues of extracting and tracking eye features from video images are addressed in this dissertation. The proposed scheme is different from conventional intrusive eye movement measuring system and can be implemented using an inexpensive personal computer. The desirable features of such a measurement system are low cost, accuracy, automated operation, and non-intrusiveness. An overall scheme is presented for which a new algorithm is forwarded for each of the function blocks in the processing system. A new corner detection algorithm is presented in which the problem of detecting corners is solved by minimizing a cost function. Each cost factor captures a desirable characteristic of the corner using both the gray level information and the geometrical structure of a corner. This approach additionally provides corner orientations and angles along with corner locations. The advantage of the new approach over the existing corner detectors is that it is able to improve the reliability of detection and localization by imposing criteria related to both the gray level data and the corner structure. The extraction of eye features is performed by using an improved method of deformable templates which are geometrically arranged to resemble the expected shape of the eye. The overall energy function is redefined to simplify the minimization process. The weights for the energy terms are selected based on the normalized value of the energy term. Thus the weighting schedule of the modified method does not demand any expert knowledge for the user. Rather than using a sequential procedure, all parameters of the template are changed simultaneously during the minimization process. This reduces not only the processing time but also the probability of the template being trapped in local minima. An efficient algorithm for real-time eye feature tracking from a sequence of eye images is developed in the dissertation. Based on a geometrical model which describes the characteristics of the eye, the measurement equations are formulated to relate suitably selected measurements to the tracking parameters. A discrete Kalman filter is then constructed for the recursive estimation of the eye features, while taking into account the measurement noise. The small processing time allows this tracking algorithm to be used in real-time applications. This tracking algorithm is suitable for an automated, non-intrusive and inexpensive system as the algorithm is capable of measuring the time profiles of the eye movements. The issue of compensating head movements during the tracking of eye movements is also discussed. An appropriate measurement model was established to describe the effects of head movements. Based on this model, a Kalman filter structure was formulated to carry out the compensation. The whole tracking scheme which cascades two Kalman filters is constructed to track the iris movement, while compensating the head movement. The presence of the eye blink is also taken into account and its detection is incorporated into the cascaded tracking scheme. The above algorithms have been integrated to design an automated, non-intrusive and inexpensive system which provides accurate time profile of eye movements tracking from video image frames.

A new artificial neural network architecture called Power Net (PWRNET) and Orthogonal Power Net (OPWRNET) has been developed. Based on the Taylor series expansion of the hyperbolic tangent function, this novel architecture can approximate multi-input multi-layer artificial networks, while requiring only a single layer of hidden nodes. This allows a compact network representation with only one layer of hidden layer weights. The resulting trained network can be expressed as a polynomial function of the input nodes. Applications which cannot be implemented with conventional artificial neural networks, due to their intractable nature, can be developed with these network architectures. The degree of nonlinearity of the network can be directly controlled by adjusting the number of hidden layer nodes, thus avoiding problems of over-fitting which restrict generalization. The learning algorithm used for adapting the network is the familiar error back propagation training algorithm. Other learning algorithms may be applied and since only one hidden layer is to be trained, the training performance of the network is expected to be comparable to or better than conventional multi-layer feed forward networks. The new architecture is explored by applying OPWRNET to classification, function approximation and interpolation problems. These applications show that the OPWRNET has comparable performance to multi-layer perceptrons. The OPWRNET was also applied to the prediction of noisy time series and the identification of nonlinear systems. The resulting trained networks, for system identification tasks, can be expressed directly as discrete nonlinear recursive polynomials. This characteristic was exploited in the development of two new neural network based nonlinear control algorithms, the Linearized Self-Tuning Controller (LSTC) and a variation of a Neural Adaptive Controller (NAC). These control algorithms are compared to a linear self-tuning controller and an artificial neural network based Inverse Model Controller. The advantages of these new controllers are discussed.

The objective of this dissertation is to develop effective algorithms for texture characterization, segmentation and labeling that operate selectively to label image textures, using the Gabor representation of signals. These representations are an analog of the spatial frequency tuning characteristics of the visual cortex cells. The Gabor function, of all spatial/spectral signal representations, provides optimal resolution between both domains. A discussion of spatial/spectral representations focuses on the Gabor function and the biological analog that exists between it and the simple cells of the striate cortex. A simulation generates examples of the use of the Gabor filter as a line detector with synthetic data. Simulations are then presented using Gabor filters for real texture characterization. The Gabor filter spatial and spectral attributes are selectively chosen based on the information from a scale-space image in order to maximize resolution of the characterization process. A variation of probabilistic relaxation that exploits the Gabor filter spatial and spectral attributes is devised, and used to force a consensus of the filter responses for texture characterization. We then perform segmentation of the image using the concept of isolation of low energy states within an image. This iterative smoothing algorithm, operating as a Gabor filter post-processing stage, depends on a line processes discontinuity threshold. Selection of the discontinuity threshold is obtained from the modes of the histogram of the relaxed Gabor filter responses using probabilistic relaxation to detect the significant modes. We test our algorithm on simple synthetic and real textures, then use a more complex natural texture image to test the entire algorithm. Limitations on textural resolution are noted, as well as for the resolution of the image segmentation process.

The Alopex algorithm is presented as a universal learning algorithm for connectionist models. It is shown that the Alopex procedure could be used efficiently as a supervised learning algorithm for such models. The algorithm is demonstrated successfully on a variety of network architectures. Such architectures include multilayer perceptrons, time-delay models, asymmetric, fully recurrent networks and memory neuron networks. The learning performance as well as the generation capability of the Alopex algorithm are compared with those of the backpropagation procedure, concerning a number of benchmark problems, and it is shown that the Alopex has specific advantages over the backpropagation. Two new architectures (gain layer schemes) are proposed for the on-line, direct adaptive control of dynamical systems using neural networks. The proposed schemes are shown to provide better dynamic response and tracking characteristics, than the other existing direct control schemes. A velocity reference scheme is introduced to improve the dynamic response of on-line learning controllers. The proposed learning algorithm and architectures are studied on three practical problems; (i) Classification of handwritten digits using Fourier Descriptors; (ii) Recognition of underwater targets from sonar returns, considering temporal dependencies of consecutive returns and (iii) On-line learning control of autonomous underwater vehicles, starting with random initial conditions. Detailed studies are conducted on the learning control applications. Effect of the network learning rate on the tracking performance and dynamic response of the system are investigated. Also, the ability of the neural network controllers to adapt to slow and sudden varying parameter disturbances and measurement noise is studied in detail.

The objective of this dissertation is to develop effective algorithms for estimating the 3-D structure of a scene and its relative motion with respect to a camera or a pair of cameras from a sequence of images acquired by the cameras, under the assumption that the relative motion of the camera is small from one frame to another. This dissertation presents an approach of computing depth maps from an image sequence, which combines the direct depth estimation method with the optical flow based method. More specifically, optical flow on and near moving edges are computed using a correlation technique. The optical flow information is then fused with the gradient information to estimate depth not only on moving edges but also in internal regions. Depth estimation is formulated as a discrete Kalman filter problem and is solved in three stages. In the prediction stage, the depth map estimated for the previous frame, together with knowledge of the camera motion, is used to predict the depth variance at each pixel in the current frame. In the estimation stage, a vector-version of Kalman filter formulation is adapted and simplified to refine the predicted depth map. The resulting estimation algorithm takes into account the information from the neighboring pixels, and thus is much more robust than the scalar-version Kalman filter implementation. In the smoothing stage, morphological filtering is applied to reduce the effect of measurement noise and fill in uncertain areas based on the error covariance information. Since the depth at each pixel is estimated locally, the algorithm presented in this paper can be implemented on a parallel computer. The performance of the presented method is assessed through simulation and experimental studies. A new approach for motion estimation from stereo image sequences is also proposed in this dissertation. First a stereo motion estimation model is derived using the direct dynamic motion estimation technique. The problem is then solved by applying a discrete Kalman filter that facilitates the use of a long stereo image sequence. Typically, major issues in such an estimation method are stereo matching, temporal matching, and noise sensitivity. In the proposed approach, owing to the use of temporal derivatives in the motion estimation model, temporal matching is not needed. The effort for stereo matching is kept to a minimum with a parallel binocular configuration. Noise smoothing is achieved by the use of a sufficiently large number of measurement points and a long sequence of stereo images. Both simulation and experimental studies have also been conducted to assess the effectiveness of the proposed approach.

In this dissertation, the digital signal processing techniques required for a 3-D sonar imaging system are examined. The achievable performance of the generated images is investigated by using a combination of theoretical analysis, computer simulation and field experiments. The system consists of a forward looking sonar, with separate projector and receiver. The projector is a line source with an 80 degrees by 1.2 degree beam pattern, which is electronically scanned within a 150 degree sector. The receiver is a multi element line array, where each transducer element has a directivity pattern that covers the full sector of view, that is 150 degrees by 80 degrees. The purpose of this sonar system is to produce three dimensional (3-D) images which display the underwater topography within the sector of view up to a range of 200 meters. The principle of operation of the proposed 3-D imaging system differs from other commonly used systems in that it is not based on the intensity of backscatter. The geometries of the targets are obtained from the delay and direction information that can be extracted from the signal backscatter. The acquired data is further processed using an approach based on sequential Fourier transforms to build the 3-D images. With careful selection of the system parameters, the generated images have sufficient quality to be used for AUV tasks such as obstacle avoidance, navigation and object classification. An approach based on a sophisticated two dimensional (2-D) autoregressive (AR) model is explored to further improve the resolution and generate images with higher quality. The real time processing requirements for image generation are evaluated, with the use of dedicated Digital Signal Processing (DSP) chips. A pipeline processing model is analyzed and developed on a selected system.