Cyber-Physical Systems (CPS) are dynamic and reactive systems interacting with processes, environment and, sometimes, humans. They are often distributed with sensors and actuators, smart, adaptive, predictive and react in real-time. Indeed, as sight for human beings, image- and video-processing pipelines are a prime source for environmental information for systems allowing them to take better decisions according to what they see. Therefore, in FitOptiVis we are developing novel methods and tools to integrate complex image and video processing pipelines. FitOptiVis aims to deliver a reference architecture for describing and optimizing quality and resource management for imaging and video pipelines in CPS both at design- and run-time. The architecture is concretized in low-power, high-performance, smart components, and in methods and tools for combined design-time and run-time multi-objective optimization and adaptation within system and environment constraints.

Cyber-Physical Systems (CPS) are systems that are in feedback with their environment, possibly with humans in the loop. They are often distributed with sensors and actuators, smart, adaptive and predictive and react in real-time. Image- and video-processing pipelines are a prime source for environmental information improving the possibilities of active, relevant feedback. In such a context, FitOptiVis aims to provide end-to-end multi-objective optimization for imaging and video pipelines of CPS, with emphasis on energy and performance, leveraging on a reference architecture, supported by low-power, high-performance, smart devices, and by methods and tools for combined design-time and run-time multi-objective optimization within system and environment constraints.

With the rapid growth of image and video data as well as the fast spread of user-generated content in social media and cloud services, it has become increasingly difficult for users to have efficient access and effective management of their digital content. In this paper we present a novel integrated open source multimedia content management and access framework, called VisualLabel, that enables smart photo services based on automated visual content analysis, annotation, search and retrieval using state of the art analysis back ends for services such as Facebook and Flickr. This paper includes detailed descriptions of the high-level architecture used in the VisualLabel framework and proof-of-concept implementations of a front-end service, along with three analysis back ends and a web client, all of which demonstrate the basic functionality provided by the framework.

Modeling and prediction of temporal sequences is central to many signal processing and machine learning applications. Prediction based on sequence history is typically performed using parametric models, such as fixed-order Markov chains ( n -grams), approximations of high-order Markov processes, such as mixed-order Markov models or mixtures of lagged bigram models, or with other machine learning techniques. This paper presents a method for sequence prediction based on sparse hyperdimensional coding of the sequence structure and describes how higher order temporal structures can be utilized in sparse coding in a balanced manner. The method is purely incremental, allowing real-time online learning and prediction with limited computational resources. Experiments with prediction of mobile phone use patterns, including the prediction of the next launched application, the next GPS location of the user, and the next artist played with the phone media player, reveal that the proposed method is able to capture the relevant variable-order structure from the sequences. In comparison with the n -grams and the mixed-order Markov models, the sparse hyperdimensional predictor clearly outperforms its peers in terms of unweighted average recall and achieves an equal level of weighted average recall as the mixed-order Markov chain but without the batch training of the mixed-order model.

We present an approach to fingerprint image enhancement that relies on detecting the fingerprint ridges based on the sign of the second directional derivative of the digital image. A facet model is used in order to approximate the derivatives at each image pixel based on the intensity values of pixels located in a certain neighborhood. The size of this neighborhood determines the scale of the image details that are preserved. We develop a selection criterion for the neighborhood size that aims to preserve minutiae details and remove smaller details from the enhanced image. Experimental results demonstrate the ability of the proposed approach to preserve a large percentage of the genuine minutiae in the enhanced image.

This paper proposes an adaptive binarization method for the document image binarization acquired by an imaging phone. The used algorithm determines the local thresholds with the information from the global trend as well as the local details. As a consequence, the proposed method is good not only for preserving the fine details of the character structure, but also for alleviating noise. The effectiveness of the proposed method is illustrated using the results obtained with a Nokia imaging phone.

A masking model originally designed for audio signals is applied to narrowband speech. The model is used to detect and remove the perceptually irrelevant simultaneously masked frequency components of a speech signal. Objective measurements have shown that the modified speech signal can be coded more efficiently than the original signal. Furthermore, it has been confirmed through perceptual evaluation that the removal of these frequency components does not cause significant degradation of the speech quality but rather, it has consistently improved the output quality of two standardized speech codecs. Thus, the proposed irrelevancy removal technique can be used at the front end of a speech coder to achieve enhanced coding efficiency.

We propose new multistage interconnection networks (MIN) for scalable parallel Viterbi decoder architectures. The architecture consists of the desired number of processing elements (PE) connected by the suggested MINs, thus allowing a tradeoff between complexity and speed. The structure of the MIN is derived first by transforming the de Bruijn interconnection-based Viterbi algorithm trellis into the equivalent trellis with a perfect shuffle interconnection, and then applying a new decomposition of the perfect shuffle operator. This results in an efficient modular system and data flow is formed by the shuffling in a local PE memory and data exchange through a fixed interconnection between PEs. We suggest several solutions for 1/n and k/n rate codes, where k denotes the number of input bits shifting into k shift registers of the encoder and, at each cycle, the encoder produces n output bits as linear combinations of certain bits in the shift registers.

This paper presents a novel and efficient coding scheme for the noise-like component in waveform interpolation speech coding. The techniques employed in the task are described in detail and the advantages gained by using them are discussed. These techniques are shown to enable perceptually transparent coding of the magnitude spectra of the noise-like component with only ten bits per frame and to offer speech quality improvements at higher bit rates. The most significant features of the proposed scheme are smoothed representation and matrix quantization of the magnitude spectra. The proposed coding techniques can be used for enhancement of speech coders based on the waveform interpolation approach.

The problem of reconstructing a one-dimensional (1D) signal from only the magnitude of its Fourier transform emerges when the phase of a signal is apparently lost or impractical to measure. Previous solutions usually employed an iterative Fourier transform (IFT) algorithm applied on a discrete approximation of a signal. The utilization of these algorithms is seriously limited by the unpredictability of their convergence. We propose several solutions to the phase retrieval problem. The first two proposed solutions uses relationships between the phase and the gain differences (GD), or gain samples (GS), in nepers. The last proposed solution uses a neural network NN for solving the problem. The NN incorporates a combination of the maximum entropy estimation algorithm with some additional nonlinear constraints. We compare our solutions by using some numerical examples. The performances under noisy conditions are also considered.

This paper presents three implementations of triple data encryption standard (3DES) algorithm on a configurable platform. Implementations are aimed at the medium access control (MAC) protocol of a multimedia-capable wireless local area network (WLAN). For this reason, very strict timing constraints as well as demands for area-efficiency are present. The MAC processing is handled by a digital signal processor (DSP) and a Xilinx Virtex field programmable gate array (FPGA) chip. The latter one is also used for the presented encryption implementations. As a result of the study, 3DES implementations with small area and reasonable throughput and, on the contrary, with large area and very high throughput are realized. Even though 3DES turns out to be quite large and resource-demanding, the implementations still leave enough chip area for the other MAC functions. Consequently, the set requirements are met and the cipher can be integrated into the system.

An efficient algorithm for singular point (core and delta) detection in fingerprint images is proposed. The algorithm is based on an efficient directional image calculation method. A novel approach for detecting the singular points using signum change curves in the directional images is also proposed. The algorithm is further optimized to achieve faster computational speed. The method was tested with a small fingerprint database and the results were compared visually to the results obtained by human experts. The results show that the algorithm is capable of detecting all the actual singular points at accurate precision with low use of computational time. The proposed algorithm outperformed a widely used algorithm (Karu and Jain, 1996) in detection accuracy and calculation speed.

Feature selection is an important part of most learning algorithms. Feature selection is used to select the most relevant features from the data. By selecting only the relevant features of the data, higher predictive accuracy can be achieved and the computational load of the classification system can be reduced. A simple method for feature selection using feedforward neural networks is presented. The method starts by using one input neuron and adds one input at time until the wanted classification accuracy has been achieved or all attributes have been chosen. The algorithm can also be used with other classification methods. Test results are given and they are promising. Our algorithm reduces the size of the feature space significantly and improves classification accuracy. Tests were performed on commonly used databases. Average classification accuracy, when using selected features, was between 79% and 100% depending on the used dataset.

In this paper, we compare the state-level and model-level tying of continuous density hidden Markov models for the multilingual acoustic modelling. Using the model-level tying technique, the number of the language dependent (LD) phoneme models of five European languages were reduced to the desired number. This tying was based on dissimilarity measure between the LD phoneme models in a bottom-up agglomerative clustering technique. This system provided 87.3% word recognition accuracy on the test set, while a comparable multilingual recognition based on the SAMPA phone inventory obtained 84.6% accuracy on the same set. The above model-level tying technique was also used for obtaining an alternative phone inventory to SAMPA such that both inventories have an equal number of phones for these five languages. The multilingual recognition systems trained for the SAMPA and alternative phone invetonries obtained 80.9% and 83.7% word accuracies on the same test set, when state-level tying was used for reducing the number of the parameters from 199k to 76k in both systems. The original LD recognition systems obtained 89.0% recognition rate with the same test set, which contained approximately 200 isolated words from SpeechDat(II) databases for each of the five languages. In this paper, the test set results are also given for the recognition systems after performing MAP language adaptation for the multilingual phone models.

Specification and Description Language (SDL) is an efficient tool for specifying and implementing complex real time systems, such as communications protocols.A Medium Access Control (MAC) protocol for a Wireless Local Area Network (WLAN) has been implemented using the language.The source C code for the protocol application is automatically generated from SDL.In this paper, the embedding of the generated protocol into a Digital Signal Processor (DSP) is described.As the high abstraction SDL model does not contain any dependencies to the target operational environment the application is adapted to DSP using tailored environment functions.These functions convert the asynchronous signalling used inside the SDL system into physical events of the DSP.In addition, as no operating system is used, the environment functions must implement other services required by the SDL generated application, such as platform clock and interrupt handling.

A signal model used in waveform interpolation (WI) coding assumes implicitly that the pitch period remains constant within a waveform to be extracted. This assumption is utilized for determining waveform boundaries and for forming the discrete Fourier transform (DFT) of the extracted waveform. Therefore, the waveforms do not describe the original signal faithfully during pitch changes even if an accurate pitch estimate is available for each time instant. Due to the distortions occurring in waveform extraction, the original signal cannot be recovered. In this paper, we discuss assumptions of the signal model and present a waveform extraction algorithm which provides asymptotically perfect reconstruction with vanishing quantization error. The algorithm uses cubic B-spline interpolation as the continuous approximation of the signal. The full benefit of the waveform extraction algorithm is obtained when an accurate pitch estimate is available.

An array processor architecture for the 2-D discrete cosine transform (DCT) based on the row-column decomposition of the 2-D DCT. The utilized 1-D DCT architectures are derived by applying the principles used to construct pipelined fast Fourier transform architectures. In general, this approach has not been used due to the irregularities found in the fast DCT algorithms. The basis of our architectural derivation is the constant geometry fast algorithms for DCT described earlier. By rescheduling the operations; an in-place algorithm can be obtained, which can be mapped onto a pipelined structure with the aid of vertical projection. In addition, a sequential matrix transposition network is described, which is based on shift-exchange units.

A parallel implementation of the 2D discrete wavelet transform on a distributed memory multiprocessor system called PARNEU is presented. The mapping has been chosen with consideration to load balancing and communication methods in order to achieve the best possible scalability and performance in transforming one single image. Detailed performance figures are included. Experimental results show that significant parallel speedup is reached with this mapping.

In this paper we have studied the use of cascade-correlation (CC) trained multilayer perceptron (MLP) networks with a maximum covariance (MC) initialization scheme for adaptive equalization of binary data bursts in a baseband digital communication system. Conventional MLP networks have been found to perform well in various adaptive equalization tasks, although they often require quite a lot of computation. Here, we have used the CC training method for MLP networks in order to be able to create a nonlinear neural network equalizer with an adaptive structure. The CC training method finds a suitable sized network for each channel response and thus decreases the amount of computation needed. To further decrease the computational load of the system, we have applied the MC initialization scheme for weight initialization, which results in faster convergence and thus decreases the amount of training needed in the network.

A mapping and realization of H.263 video encoder for video conferencing applications are given using our DSP based multiprocessor system, called PAKNEU. PARNEU system is built for computationally intensive applications like image and video processing as well as soft computing applications. PARNEU has flexible communication architecture and thus it allows different mapping possibilities. The presented data parallel mapping has low communication and memory requirements, which allows encoding of any of the five standard H.263 picture formats. With a prototype system using four ADSP-21062 DSPs, a real-time encoding is achieved with QCIF sized picture. Phase times for each step of H.263 encoder are presented.

In this paper a real-time video encoder following ITU-T H.263 recommendation is described and relative computational loads of various encoding phases are evaluated. The current implementation shows that QCIF size (176/spl times/144) images can be coded in real-time by TMS320C6201 when computationally light motion estimation routines and basic H.263 coding mode is used. The real-time performance can be achieved using C-code and optimizing some key functions like SAD (sum of absolute difference) with linear assembler. In addition, careful memory management design for program code and application data is required. With the presented coding scheme, up to 29 fps can be achieved.

In parallel architectures for discrete trigonometric transforms, the number of processing elements is typically dependent on the transform size. Scalable architectures can be constructed with a partial column approach where the computation is performed iteratively with less number of processing elements. This approach results in a need for complex data reordering for realizing the interconnections between the processing columns. In this paper, such interconnection networks performing temporal and spatial reordering are proposed. These networks realize the data reordering found in constant geometry radix-2/sup r/ algorithms, which exist, e.g., for discrete Fourier, sine, cosine, and Hartley transforms. A general decomposition of stride by 2/sup r/ permutation is shown with corresponding network implementations. Furthermore, modifications to support mixed-size and 2-D transforms are discussed.

A distance transform (DT) converts a binary image consisting of foreground (feature) and background (non-feature) pixels into a gray level image where each pixel contains the distance from the corresponding pixel to the nearest foreground pixel. The computation of the exact Euclidean DT is computationally complex task and, therefore, approximations are typically utilized. An area-efficient architecture for computing a DT approximation is presented. The architecture utilizes order-based encoded distance representation allowing simple bitwise operations to be used for determining the distance to the nearest foreground pixel in the constrained neighborhood. Tabulated distance values are used thus cumulative errors are avoided. Due to the simple operations real-time operation can be expected.

This paper compares two noise compensation techniques for increasing recognition performance on tone corrupted speech from payphones. An analysis of payphone speech is made which identifies two harmful processes signalling tones and increased background noise. The techniques of spectral subtraction and parallel model combination are compared on this task and are shown to give some improvement. Further robustness is demonstrated using RASTA-CTM speech features.

In recent work, the all-pass transform (APT) was proposed as the basis of a speaker adaptation scheme intended for use with a large vocabulary speech recognition system. It was shown that APT-based adaptation reduces to a linear transformation of cepstral means, much like the better known maximum likelihood linear regression (MLLR), but is specified by far fewer free parameters. Due to its linearity, APT-based adaptation can be used in conjunction with speaker-adapted training (SAT), an algorithm for performing maximum likelihood estimation of the parameters of an HMM when speaker adaptation is to be employed during both training and test. In this work, we propose a refinement of SAT called single-pass adapted training (SPAT) which achieves the same improvement in system performance as SAT but requires much less computation for HMM training. In a set of speech recognition experiments conducted on the Switchboard Corpus, we report a word error rate reduction of 5.3% absolute using a single, global APT.

The support for multimedia services in wireless access networks is challenging to implement, due to an unreliable and bandwidth-limited wireless medium. In addition, the widely used applications and protocols generally lack the support for quality-of-service (QoS) parameters. This paper presents the architecture of a multimedia wireless LAN terminal. The terminal is implemented using a custom network demonstrator platform connected to a Windows NT workstation. The system provides a separate management plane for configuring the service parameters of a proprietary wireless medium access control (MAC) protocol. Also, native applications that are capable of accessing the MAC QoS parameters directly are supported.

The software development and optimization of the communication is presented for an experimental scalable multiprocessor system purposed for multimedia applications. For convenient application mappings, presented C-code primitives hide the communication topology and allow the system to be expanded without needs to rewrite programs. For real-time applications, fast hardware communication need a strong support from software to prevent large latency times. Measured latency times and obtained throughput in communication are compared between user level application and maximum available hardware performance.

This paper presents a medium access control (MAC) protocol called TUTMAC for a new wireless local area network (TUTWLAN). The design objective has been to develop a simple, multimedia service capable protocol that provides sufficient medium utilisation efficiency and guarantees QoS (quality of service) parameters. The developed system utilises a centralised (base station controlled) network architecture. A limited number of portable stations can be associated with the same base station, i.e. in the same TUTWLAN cell. TUTMAC is connection oriented: the bandwidth is allocated deploying constant bit-rate TDMA based data channels that are reserved by exchanging short control messages. The connection parameters can be dynamically altered during the data exchange session. Currently, a TUTWLAN prototype is being developed comprising both TUTMAC software and platform hardware modules. The prototype will support up to eight simultaneous data-transfer connections each having 64 to 512 kbit/s data transmission bandwidth.

This paper presents a security scheme for a medium access control protocol in a new wireless local area network TUTWLAN (Tampere University of Technology WLAN). The design objective has been to develop a security scheme that will be scalable for various needs and offer high security for demanding applications. The designed security scheme provides both privacy of wireless data communications and the authenticity of communicating parties. Our authentication scheme allows also the communicating entities to establish a shared secret key for secure communication session. Data security schemes have also been introduced. There are three optional data security modes that offer flexible ciphering and data security level.

Although neural networks have already found many successful applications in speech signal processing, it is difficult to compare their performance with that of the traditional methods objectively. We address this problem by using the predictive minimum description length (PMDL) principle to select the optimal network size in modeling speech signals.

This paper presents two parallel realizations of sparse distributed memory (SDM) on a tree-shaped computer. The original model of SDM is introduced in terms of generalized computer memory and artificial neural networks (ANNs). For parallellization purposes, addressing, storage and retrieval operations are explained in detail. Some existing implementations in various computing platforms are considered before introducing the tree-shaped parallel computer, TUTNC (Tampere University of Technology Neural Computer). Two mappings are given, each utilizing parallelism with different granularities, and compared in terms of measured execution time, task partitioning and load balancing. Performance estimates are given for a larger system. The results show that SDM can be well parallelized in TUTNC. © 1997 John Wiley & Sons, Ltd.

Nonlinear time series modeling with a multilayer perceptron network is presented. An important aspect of this modeling is the model selection, i.e., the problem of determining the size as well as the complexity of the model. To overcome this problem we apply the predictive minimum description length (PMDL) principle as a minimization criterion. In the neural network scheme it means minimizing the number of input and hidden units. Three time series modeling experiments are used to examine the usefulness of the PMDL model selection scheme. A comparison with the widely used cross-validation technique is also presented. In our experiments the PMDL scheme and the cross-validation scheme yield similar results in terms of model complexity. However, the PMDL method was found to be two times faster to compute. This is significant improvement since model selection in general is very time consuming.

Usually the training of a multilayer perceptron network starts by initializing the network weights with small random values, and then the weight adjustment is carried out by using an iterative gradient descent-based optimization routine called backpropagation training. If the random initial weights happen to be far from a good solution or they are near a poor local optimum, the training will take a lot of time since many iteration steps are required. Furthermore, it is very possible that the network will not converge to an adequate solution at all. On the other hand, if the initial weights are close to a good solution the training will be much faster and the possibility of obtaining adequate convergence increases. In this paper a new method for initializing the weights is presented. The method is based on the orthogonal least squares algorithm. The simulation results obtained with the proposed initialization method show a considerable improvement in training compared to the randomly initialized networks. In light of practical experiments, the proposed method has proven to be fast and useful for initializing the network weights.

An optimization tool for neural network architecture selection is presented. The main aim of the optimization tool is to reduce the size and complexity of the network and use the least number of weights and nodes for modeling and predictions on nonlinear time series. The problem of selecting the number of input and hidden nodes for modeling nodes is studied by the predictive minimum description length (MDL) principle. The authors discuss comparatively the performance of neural networks and conventional methods in predicting nonlinear time series. The neural network is found to yield better predictions than an optimum ARMA (autoregressive moving average) model.