In this thesis, estimation of motion from an image sequence is investigated. The emphasis is on the novel use of motion model for describing two dimensional motion. Special attention is directed towards general motion models which are not restricted to translational motion. In contrast to translational motion, the 2-D motion is described by the model using motion parameters. There are two major areas which can benefit from the study of general motion model. The first one is image sequence processing and compression. In this context, the use of motion model provides a more compact description of the motion information because the model can be applied to a larger area. The second area is computer vision. The general motion parameters provide clues to the understanding of the environment. This offers a simpler alternative to techniques such as optical flow analysis. A direct approach is adopted here to estimate the motion parameters directly from an image sequence. This has the advantage of avoiding the error caused by the estimation of optical flow. A differential method has been developed for the purpose. This is applied in conjunction with a multi-resolution scheme. An initial estimate is obtained by applying the algorithm to a low resolution image. The initial estimate is then refined by applying the algorithm to image of higher resolutions. In this way, even severe motion can be estimated with high resolution. However, the algorithm is unable to cope with the situation of multiple moving objects, mainly because of the least square estimator used. A second algorithm, inspired by the Hough transform, is therefore developed to estimate the motion parameters of multiple objects. By formulating the problem as an optimization problem, the Hough transform is computed only implicitly. This drastically reduces the computational requirement as compared with the Hough transform. The criterion used in optimization is a measure of the degree of match between two images. It has been shown that the measure is a well behaving function in the vicinity of the motion parameter vectors describing the motion of the objects, depending on the smoothness of the images. Therefore, smoothing an image has the effect of allowing longer range motion to be estimated. Segmentation of the image according to motion is achieved at the same time. The ability to estimate general motion in the situation of multiple moving objects represents a major step forward in 2-D motion estimation. Finally, the application of motion compensation to the problem of frame rate conversion is considered. The handling of the covered and uncovered background has been investigated. A new algorithm to obtain a pixel value for the pixels in those areas is introduced. Unlike published algorithms, the background is not assumed stationary. This presents a major obstacle which requires the study of occlusion in the image. During the research, the art of motion estimation hcis been advanced from simple motion vector estimation to a more descriptive level: The ability to point out that a certain area in an image is undergoing a zooming operation is one example. Only low level information such as image gradient and intensity function is used. In many different situations, problems are caused by the lack of higher level information. This seems to suggest that general motion estimation is much more than using a general motion model and developing an algorithm to estimate the parameters. To advance further the state of the art of general motion estimation, it is believed that future research effort should focus on higher level aspects of motion understanding.

Learning the count list (one, two, three, …) is a critical stepping-stone for the acquisition of number concepts. Most research about counting, however, is done in the behavioral domain, and little is known about the neural representations underlying counting sequences. Here, we test the hypothesis that transitional knowledge within a counting sequence exist both at sensory and conceptual (ordinal and magnitude) levels. To test this hypothesis, we employed a passive-listening violation-to-expectation fMRI paradigm where adult participants heard auditory count sequences that were correct (4 5 6 7) or violated at the end (4 5 6 8; consecutiveness) and, orthogonally, that were ordered or unordered (orderedness). Another orthogonal dimension was the manipulation of sensory sequence violation where the voice speaking the numbers was consistent throughout the trial or could change on the last number (voice identity). This 2x2x2 factorial design was analyzed using univariate and multivariate pattern analyses. Three clusters in the right fronto-parietal network (BA44, BA46, and IPS) showed greater neural response to violations to orderedness. Of the three clusters, the anterior IFG (BA46) demonstrated the encoding of consecutiveness. Interestingly, the bilateral STG, which showed a robust effect to violations in voice identity, also demonstrated the encoding of consecutiveness. These results indicate that a right-lateralized fronto-parietal network activity can differentiate between a count list and random numbers, while BA46 and bilateral STG respond specifically to violations of the count sequence, suggesting specific mechanisms in the brain for processing consecutive numbers in both the perceptual and cognitive levels.

The physical limitations of uniprocessors and the real-time requirements of numerous practical applications have made parallel processing an essential technology in military, industry and scientific research. In this dissertation, we investigate parallelizations of three practical applications using three parallel machine models. The algorithms are: Finitely inductive (FI) sequence processing is a pattern recognition technique used in many fields. We first propose four parallel FI algorithms on the EREW PRAM. The time complexity of the parallel factoring and following by bucket packing is O(sk^2 n/p), and they are optimal under some conditions. The parallel factoring and following by hashing requires O(sk^2 n/p) time when uniform hash functions are used and log(p) ≤ k n/p and pm ≈ n. Their speedup is proportional to the number processors used. For these results, s is the number of levels, k is the size of the antecedents and n is the length of the input sequence and p is the number of processors. We also describe algorithms for raster/vector conversion based on the scan model to handle block-like connected components of arbitrary geometrical shapes with multi-level nested dough nuts for the IES (image exploitation system). Both the parallel raster-to-vector algorithm and parallel vector-to-raster algorithm require O(log(n2)) or O(log2(n2)) time (depending on the sorting algorithms used) for images of size n2 using p = n2 processors. Not only is the DWT (discrete wavelet transforms) useful in data compression, but also has it potentials in signal processing, image processing, and graphics. Therefore, it is of great importance to investigate efficient parallelizations of the wavelet transforms. The time complexity of the parallel forward DWT on the parallel virtual machine with linear processor organization is O(((so+s1)mn)/p), where s0 and s1 are the lengths of the filters and p is the number of processors used. The time complexity of the inverse DWT is also O(((so+s1)mn)/p). If the processors are organized as a 2D array with PrawPcol processors, both the interleaved parallel DWT and IDWT have the time complexity of O(((so+s1)mn)/ProwPcol). We have parallelized three applications and achieved optimality or best-possible performances for each of the three applications over each of the chosen machine models. Future research will involve continued examination of parallel architectures for implementation of practical problems.

No description available.

This work investigates the self-organizing representation of temporal data in prototype-based neural networks. Extensions of the supervised learning vector quantization (LVQ) and the unsupervised self-organizing map (SOM) are considered in detail. The principle of Hebbian learning through prototypes yields compact data models that can be easily interpreted by similarity reasoning. In order to obtain a robust prototype dynamic, LVQ is extended by neighborhood cooperation between neurons to prevent a strong dependence on the initial prototype locations. Additionally, implementations of more general, adaptive metrics are studied with a particular focus on the built-in detection of data attributes involved for a given classifcation task. For unsupervised sequence processing, two modifcations of SOM are pursued: the SOM for structured data (SOMSD) realizing an efficient back-reference to the previous best matching neuron in a triangular low-dimensional neural lattice, and the merge SOM (MSOM) expressing the temporal context as a fractal combination of the previously most active neuron and its context. The first SOMSD extension tackles data dimension reduction and planar visualization, the second MSOM is designed for obtaining higher quantization accuracy. The supplied experiments underline the data modeling quality of the presented methods.

This master thesis deals with the implementation of various types of recurrent neural networks via programming language lua using torch library. It focuses on finding optimal strategy for training recurrent neural networks and also tries to minimize the duration of the training. Furthermore various types of regularization techniques are investigated and implemented into the recurrent neural network architecture. Implemented recurrent neural networks are compared on the speech recognition task using AMI dataset, where they model the acustic information. Their performance is also compared to standard feedforward neural network. Best results are achieved using BLSTM architecture. The recurrent neural network are also trained via CTC objective function on the TIMIT dataset. Best result is again achieved using BLSTM architecture.

This thesis proposes tasks of exercises for mentioned course and design and creation of automated evaluation system for these exercises. This thesis focuses on discussion and exemplary solutions of possible tasks of each exercise and description of created automated evaluation system. For evaluation program are made tests with chosen special data sets, which will prove it’s functionality in general data sets.

This thesis deals with lossless and losing methods of data compressions and their possible applications in the measurement engineering. In the first part of the thesis there is a theoretical elaboration which informs the reader about the basic terminology, the reasons of data compression, the usage of data compression in standard practice and the division of compression algorithms. The practical part of thesis deals with the realization of the compress algorithms in Matlab and LabWindows/CVI.