3. Learning and intelligent image and signal analysis (LIISA)

Abstract

Results

Project information

Participants

Project dates

Project volume

Project manager

Publications

Iivarinen, J., Visa, A., Shape Recognition of Irregular Objects, in D. P. Casasent (Ed.), Intelligent Robots and Computer Vision XV: Algorithms, Techniques, Active vision, and Materials Handling, Proc. SPIE 2904, pp. 25-32, 1996.

Iivarinen, J., Rauhamaa, J., Visa, A., An Adaptive Approach to Segmentation of Surface Defects, Report A34 , Helsinki University of Technology, Faculty of Information Technology,  Laboratory of Computer and Information Science, March, 15 p., 1996.

Iivarinen, J., Rauhamaa, J., Visa, A., An Adaptive Two-Stage Approach to Classification of Surface Defects, in Proc. of the 10th Scandinavian Conference on Image Analysis, Lappeenranta, Finland, June 9-11, pp. 317-322, 1997.

Iivarinen, J., Peura, M., Särelä, J., Visa, A., Comparison of Combined Shape Descriptors for Irregular Objects, in Proc. of the 8th British Machine Vision Conference, University of Essex, UK, September 8-11, vol. 2, pp. 430-439, 1997.

Iivarinen, J., Visa, A., An Adaptive Texture and Shape Defect Classification, in Proc. of 14th IAPR International Conference on Pattern Recognition, Brisbane, Australia, August 16-20, vol. I, pp. 117-122, 1998.

Iivarinen, J., Visa, A., Unsupervised Image Segmentation with The Self-Organizing Map and Statistical Methods, in D. P. Casasent (Ed.), Intelligent Robots and Computer Vision XVII: Algorithms, Techniques, and Active vision, Proc. SPIE 3522, pp. 516-526, 1998.

Iivarinen, J., Rauhamaa, J., Surface Inspection of Web Materials Using The Self-Organizing Map, in D. P. Casasent (Ed.), Intelligent Robots and Computer Vision XVII: Algorithms, Techniques, and Active vision, Proc. SPIE 3522, pp. 96-103, 1998.

Heikkinen, K., Vuorimaa, P., Hardware Architecture of the SOM for Real-Time Applications, in Proc. Of the 5^th European Congress on Intelligent Techniques and Soft Computing, vol. III, pp 2495-2499, Aachen, Germany, September 8-11, 1997.

Heikkinen K., Vuorimaa, P., Computation of two texture features in hardware, in proc. of the International Conference on Acoustics, Speech, and Signal Processing, pp. xxx-xxx, Phoenix, Arizona, USA, March 15-19, 1999.

Holmström, L., Koistinen, P., Laaksonen, J., Oja, E., Neural Network and Statistical Perspectives of Classification, in Proc. 13^th International Conference on Pattern Recognition, Vienna, Austria, August 25-30, vol IV, pp 286-290, 1996.

Holmström, L., Koistinen, P., Laaksonen, J., Oja, E., Comparison of Neural and Statistical Classifiers – Theory and two Practice, University of Helsinki, Rolf Nevanlinna Institute, Research Reports A13, 37 p. 1996.

Holmström, L., Koistinen, P., Laaksonen, J., Oja, E., Neural and Statistical Classifiers – Taxonomy and two Case Studies, IEEE Transactions on Neural Networks, 8:5-17, 1997.

Laaksonen, J., Oja, E., Classification with Learning k-Nearest Neighbors, in Proc. International Conference on Neural Networks (ICNN’96), Washington DC, USA, June 3-6, vol. 3, pp. 1480-1483, 1996.

Laaksonen, J., Oja, E., Subspace Dimension Selection and Averaged Learning Subspace Method in Handwritten Digit Classification, in Proc. International Conference on Artificial Neural Networks (ICANN’96), Bochum, Germany, July 16-19, pp. 227-232, 1996.

3.2 Intelligent signal analysis and the required component technology

Abstract

Results

1. Classification

There are special clinics where accurate clinical methods can be used to measure the maximal oxygen uptake, but the measurements are expensive and the tests are time-consuming. The need for an inexpensive and reasonably accurate system for measuring aerobic fitness is therefore obvious. The purpose of this study was to create a method for measuring aerobic fitness from a relatively short time series of data (R-R intervals) without extensive physiological tests. The material for this study was obtained from the Merikoski Institute of Health Research and Rehabilitation, Oulu, Finland.

Fitness approximation

Figure 1. Block diagram of the method.

The Levenberg-Marguardt (LM) optimising algorithm was used to train the ANN. The LM method was superior to the other training algorithms, since the LM procedure produced the most optimal results for ANN simulation. The training results were verified with a cross-checking method. The results are based on simulations of the ANN structure created. The correlation of the simulation values compared to the accurate clinical measurements was 0.97 for the training material and 0.96 for the testing material.

Using the neural network developed here, it is possible to define the maximal oxygen uptake from a short R-R interval measurement without a testing protocol with high correlation (0.96) and high accuracy (mean error 6.5%). This is a new and easy way to determine and control personal aerobic fitness, even at home.

A prototype of the model was also constructed. The estimation program operates in a MATLAB environment. The code includes signal filtering, feature extraction and, ultimately, neurocalculation. The program was tested with several test persons, and the results were promising. The aerobic fitness approximation method will be implemented in a Windows 95 program. The purpose of the program is to validate the novel method in field tests in several research clinics.

A multi-structure solution for the context-sensitive feature selection problem

Figure 2. Operating principle of the fuzzy preclassifier.

Parallel-structured neural networks for classification
The presented multistructure solution is a modification of the network-boosting algorithm. The input patterns are statistically analysed and clustered. A preclassifier is then constructed based on the statistical clusters. The purpose of the preclassifier is to classify roughly similar patterns into different classes and, in that way, to select the final classification structure to be used. The use of several neural structures for the classification problem preserves the local characteristics of the features. In this way, the problem of context-sensitive features is solved and each structure has its own characteristics and weight space for calculating the final output of the hybrid system.

The training material is divided into clusters with the optimal principle. Some overlap is used in classification. The reason for this is that the preclassifier cannot have perfect accuracy in absolute terms, but the failed classes are, with high probability, in the cluster next to the proposed structure. This overlap requires the pattern classes to be sorted with some similarity criteria.

The final form of the each structure is created using a genetic-like algorithm. The starting weights are randomly generated, as is the network layer size. The network is trained with the given parameters and attributes. The outcome of the training is evaluated with cross-checking. The best results are recorded, including the size of the network and the weight information. This looping is done several thousand times to find the optimal size of the structure as well as the global minimum of the weight space. It should be pointed out that a genetic algorithm of this kind requires very intensive calculating power.

Human aerobic fitness classification

The accuracy of the fuzzy preclassifier was 80.9 %, indicating low-level fitness samples, and 86.2 %, indicating samples of good fitness level. It should be pointed out that the classified samples were all correctly classified. The 20 % of the samples were not recognised as examples of a good or a poor fitness level.

For the sake of comparison, all the samples were classified using the M structure, which was implemented conventionally using a single neural network structure. The accuracy of the single structure classifier was 50 %.

When the whole chain of the proposed method was used, the following results were obtained. The accuracy of the S network was 89 %, that of the M network 45 % and that of the L network 70 %. This overall accuracy of the neuro-fuzzy system was 63.2 %. The accuracy was 13.2 % better than that obtained using a conventional neural network classifier.

The results were very promising. With this technique, it is possible to divide even a very large set of training data into smaller parts for different neural structures. This division improves accuracy and reduces the time needed for neural network training. In the test case, the resulting accuracy of 63.2 % was not optimal in view of the classification application. However, it is the best result that can be achieved using the biological features of the person in a resting state.

The generality of the method proposed is due to its modularity. In the literature, modular networks have been presented, but their structure selection has not been based on a fuzzy preclassifier. In our approach, modularity is created using statistical analysis of features and fuzzy analysis. The classification space is divided into smaller parts with significant similarity. A symmetric structure is then created, which requires measurement of the distance between the used classes. The classes and structures are sorted using the distances measured. In this way, the aerobic fitness classification obtained by using neuro-fuzzy approach and a modular network structure can be generalised to other real-world classification problems and applications.

Intensifying NN learning with dimension reduction

Traditional PCA can only be used for linear data, and because the physiological functions of a human being are highly non-linear, the data have to be divided to internally linear groups. After this, PCA can be performed independently for each group.

Another possibility is to use non-linear principal components. They can be constructed with neural networks. An NLPCA network is an auto-associative network with three hidden layers: a mapping layer, a bottleneck layer and a de-mapping layer. The bottleneck layer will produce an effective but lower dimensional presentation of the original data.

The study population consisted of 237 healthy men and women aged 15-65 years, of whom 203 subjects were used as the training set and 34 subjects as an independent test set.

Both non-linear and linear PCA reduced the dimensions from 18 original variables to 8 principal components. Maximal oxygen uptake, which was normalised for the subject’s weight (ml/kg/min), was approximated with three different data sets (18 original parameters, 8 linear PCs, 8 NLPCs). MLP with the Levenberg-Marquardt optimisation algorithm was used in the approximation. The optimal learning result was selected by minimising the RMS error for the test set.

The best result was achieved with 8 non-linear principal components, and the correlations between the target and predicted values were 0.95 for the training set and 0.94 for the test set. The mean errors were 7% for the training set and 9% for the test set. The results of the three input sets did not differ much from each other in accuracy, but the training time was over twofold longer when the original variables were used instead of the non-linear principal components. On the other hand, no significant amount of information was lost due to the dimension reduction.

2. Synchronisation

A neural network-based synchroniser for a CDMA spread spectrum system was developed. The goal was to realise a fast-channel delay estimator when one or two signals are transmitted through the channel. The performance of the multilayer neural network as a channel delay estimator was studied with simulations. The results were compared to those of an ML estimator derived in the work. The neural network used in the simulations was a multilayer perceptron with a back-propagation learning algorithm and a complex-valued Kalman filter algorithm.

New digital filtering methods are being developed for filtering audio interference from the speech signal. A method based on timing information of the interference is used to control the filtering. The algorithms are designed to be adaptable, enabling the filter to learn from interference on the basis of samples. The filter is also able to detect speech and non-speech segments of the signal, and hence to control dynamic adaptation during the filtering process. A coherent subtraction method and an adaptive line enhancer filter bank approach were developed in 1997. Listening tests have been used to evaluate the results.

Neural network-based implementation of a synchronising circuit in CDMA systems

In this work, only one-user and two-user systems were examined. Thus, the duration of simulations in some cases was several weeks. One task was to examine how simulations could be shortened without sacrificing the quality of the results.

To reduce the complexity of the research problem, the effects of the complexity parameters were examined one by one. The most important parameters were: the phase shift of the carrier wave between two sent signals and the variation of relative signal strength in the case of two users. The near-far problem, where strong signals interfere with the reception of weaker signals, was especially considered.

Results with the one-user system

Envelope detectors were added to the system model in contradiction to the ML-estimator. The network was trained with noiseless training material. With noisy test signals, the performance of the network was inferior to that of a non-optimal estimator based on maximum search. The network learned to estimate the training signals accurately, but failed to do so with noisy signals. With noiseless test signals, the performance of the network was better than that of the non-optimal estimator.

Results with the two-user system

These parameters were used to generate the training data: the relative intensity of the signals (NFR in decibels), the delays of the users A and B, and the phase angles of the carrier waves of the users A and B.

The network did not learn to estimate the channel delay well in the noisy two-user system. In addition, the simulation times and the size of the training material increase exponentially with the number of users in the system. The results can be improved in simple cases by post-integrating the estimates, but in more complicated cases this has no effect. The present results indicate that the multi-user synchronisation problem should be divided into smaller sub-problems, some of which could be solved by a neural network.

Adaptive filtering of cyclostationary interference from a speech signal

The interference originates the TDMA system’s frame rate. The amplitude modulation of the sending signal causes electromagnetic interference with audio-signals in the electrical circuits near the transmitter.

The interference has two important properties: timing effects of GSM data and external effects, e.g. movement of the receiver in relation to the terminal changes the amplitude of the interference. The timing properties of GSM data can be estimated from a speech signal by utilising knowledge of the known timing information of the interference. External effects cannot be predicted and the algorithm should therefore be adaptive.

Figure 3. Original (top) and filtered signal.

Results with speech filtering

Based on the literature and previous experience, two different methods were developed: a coherent subtraction method and an adaptive line enhancer filter bank approach.

Listening tests were used to evaluate the results. With both methods, interference in a speech signal was reduced to a nearly acceptable level. A subtraction method in the time domain gave better speech quality, because filtering in the frequency domain also modifies the spectrum of the filtered speech signal.

A coherent subtraction method required interpolation and decimation to increase the sampling rate from 8 kHz to 9.1 kHz: otherwise, the interference drifted in relation to the sampling rate. Interpolation and decimation were overly heavy operations for implementation, and some work was therefore done to model the drifting effect of interference.

The time domain method is being further developed by creating an parameter model of the known behaviour of the interference and incorporating the adaptive features that have been developed.

3. VLSI implementation

Neuron architectures

During the project period a fully analogue cell for a Herault-Jutten network was designed and fabricated. A commercial 1.2 um CMOS technology was used, and the total current consumption of the chip that implements a two-input-two-output Herault-Jutten network is 10 uA. The chip is functional, and most of the cells operate as intended. However, some offset voltage problems associated with the weight cell have been encountered, and they prevent use in actual applications at the moment. Fig. 4 shows the layout of the integrated Herault-Jutten circuit.

The usability of a Herault-Jutten network in signal and noise separation was studied in simulations. Two-channel ECG signal recordings were made using different electrode orientations with respect to the heart to obtain test signals for the simulations. No noticeable signal separation or improvement in signal quality was seen in the simulations. However, by using principal component analysis, marked improvement was achieved, which is demonstrated in Fig. 5.

Figure 4. The Herault-Jutten chip.

Figure 5. ECG signal quality improvement by principal component analysis.

Project information

Participants

The work was carried out in cooperation with the following industrial companies:

• Fincitec Oy (Yrjö Mäkelä)
• Polar Electro Oy (Ilkka Heikkilä)
• Nokia Mobile Phones (Jorma Lilleberg)
• Nokia Cellular Systems (Ilkka Keskitalo)

Project dates

Project volume

Project manager

Publications

Kauko Väinämö, Seppo Nissilä, Timo Mäkikallio, Mikko Tulppo, Juha Röning, “Artificial Neural Networks for aerobic fitness Approximation”, International Conference on Neural Networks (ICNN ‘96), Washington DC, USA, June 3-6, 1996.

Kauko Väinämö, Juha Röning, ”An artificial neural network -based rotation correction method for 3D-measurement using a single perspective view”, International Conference on Quality Control by Artificial Vision (QCAV ’97), Le Creusot, France, May 28-30, 1997.

Kauko Väinämö, Timo Mäkikallio, Mikko Tulppo, Juha Röning,”A Neuro-Fuzzy Approach to Aerobic Fitness Classification: a multistructure solution to the context-sensitive feature selection problem”, International Joint Conference on Neural Networks (IJCNN ‘98), Anchorage, USA, May 4-9, 1998.

Satu Lipponen, Kauko Väinämö, Timo Mäkikallio, Mikko Tulppo, Juha Röning, “Approximating aerobic fitness with neural networks”, Finsig, Pori, 1997.

Satu Lipponen, Timo Mäkikallio, Mikko Tulppo, Juha Röning, “Finding structure in fitness data”, Practical application in Knowledge Discovery and Data Mining (PADD98), London, UK, March 25-27, 1998.

Lasanen, K., J. Tervaluoto, J., A. Ruha, A., Röning, J., A Structure for Extending the Linear Input Voltage Range of a Differential Input Stage, The 5th International Conference on Electronics, Circuits and Systems (ICECS98), 7-10 September, 1998, Lissabon, Portugal, pp. 355-358.

Väinämö, K.,Neural Networks for Human Aerobic Fitness Approximation, University of Oulu, Department of Electrical Engineering, Diploma Thesis, 1996, 89 pp.

Mannerkoski, J., Neural Network based Synchronisation in a CDMA System, University of Oulu, Department of Electrical Engineering, Diploma Thesis, 1996, 1995, 60 pp. + appendices.

Mäkelä J. Signal Subspace Tracking in Multiuser Delay Estimation. Diploma Thesis (in Finnish), University of Oulu, 1997, 62 pp.

Lasanen, K.: Analog CMOS Implementation of an Adaptive Neuron Cell, Diploma Thesis (in Finnish), University of Oulu, 1997, 49 pp.

Tervaluoto, J.: Micropower Range Neuron Integration Using CMOS-technology, Diploma Thesis (in Finnish), University of Oulu, 1997, 71 pp.

3.3 Neuro-Fuzzy systems

Abstract

• Modelling of measurements signals
• Processing of unreliable information
• Processing of video and audio signals (including hardware implementations)

Results

1. Modelling of measurements signals

Signal processing of a quality sensor with neural networks

The values of the reflection spectrum channels strongly depend on each other. To compress the information of the channels, principal component analysis (PCA) was used. With PCA the original information can be expressed as a small number of variables. A model based on either principal component regression (PCR) or neural network (radial basis function, RBF) was established between the new variables and the values of the quality units. The t-statistic was used to select the most significant principal components for the model. As a result of the research it was concluded that it is possible to make prediction models for some of the quality units. With the model it is possible to measure the values of quality units continuously. During the project there was relatively little data available and the neural network models predicted the values of quality units only slightly better than the models containing simple linear structures. Prediction errors were generally equally good with the PCR and the neural networks.

Applying neural networks to on-line pulp quality measurements

A multilayer perceptron (MLP) network was the first method to try. A basic MLP network with one hidden layer, trained with Levenberg-Marguard algorithm, works properly most of the time but it sometimes gives very strange results when the test data is clearly different from the training data. It also usually saturates at a certain level and thus can not estimate the quality levels outside the range of the training data. Training the MLP network requires a lot of data, which is not usually possible during the installation and fine-tuning phase of the device. Further development of the algorithms or the network structures was not considered necessary since different methods seemed more promising. Self-organizing maps (SOM) were used to visualise the behaviour of the device in different conditions and to classify which kind of wood the pulp was made of. Learning vector quantization was also used in classification. We discovered, however, that there were two different kinds of data: in one group the data is strongly dependent on the different woods and another group is not. A first order Sugeno-style fuzzy inference system (FIS) has been the best method so far. FIS has been effective, especially using the wood-dependent data. It is also simple; a set of linear models for different situations connects smoothly. Fuzzy rules are easy to understand and there are no too many parameters to tune. One of the main problems was to identify the reason why the device behaves in a totally different way with some data sets. The device under study was found to be relatively linear and the new methods examined gave no distinct improvement in the performance.

Neural network modelling of the pumping process

In the case of the pumping process, the forward problem can be defined with characteristic curves.  In this work the modelling of the characteristic curves and solving the inverse problem with the MLP neural network was studied. The MLP neural network is a flexible method for nonlinear modelling. This work includes a review describing theory, characteristics, training and selection of model complexity of the MLP network. There are several methods to produce inverse MLP network models. This study included a review describing the most common methods and their pros and cons.

The modelling of the characteristic curves was tested with three different data sets using advanced methods for MLP network training and model complexity selection. Four different methods were tested for solving the inverse problem.

2. Processing of unreliable information

Applying neural network algorithms to estimate the position of the fabric edge

Signal feature extraction using fuzzy gates

Prediction model for the lime kiln process / SOM training and visualizing tool

Adaptive equalization of binary data bursts

In our simulations, both MLP network and adaptive clustering method outperformed the linear equalizer, when bit error rates (BERs) were considered. Our simulation cases included simulations in both fixed and altering channels. The adaptive clustering method achieved the smallest BER. However, the BER of the MLP network was only slightly worse. The computational load of the adaptive clustering method and linear equalizer was very small compared to the MLP network. The disadvantage of the MLP network was that the training did not always converge to an adequate solution. The main reason for this was the random initialization of the network weights. To avoid this problem, an efficient weight initialization method, such as the maximum covariance (MC) initialization scheme, could he used. By applying MC initialization, the MLP network converges much faster than a randomly initialized network. Due to this the MLP network with MC initialization achieves better BERs than the randomly initialized network, if they are trained with the same amount of training epochs. Conversely, we can say that the BERs achieved by the randomly initialized MLP network can be reached much faster and/or with a smaller network by applying MC initialization. Since the MC initialization scheme is also computationally relatively light, we can actually significantly decrease the amount of computation needed in the network by applying MC initialization.

Chemical agent detection

The current chemical warfare agent recognition capability of the M90 is very good, but problems occur when multiple agents are mixed together. The neural network is taught to recognise pure gases and thus it cannot deduce the components of some gas mixtures, especially when the mixture components react with each other and create a totally new gas. Because of this, new methods were created to enable recognition of gas mixtures.

Anaesthetic machine fault diagnosis

Fuzzy classifier in defect diagnosis of solder joints

The inspection results of nearly 900 circuit boards have been analysed to distinguish the essential features of defective and non-defective solder joints. The characters detected were then expressed by linguistic rules. In addition, the effect of the prevailing average amount of solder in the joint, which is one of the major factors affecting quality, has been considered. The simulation results of the behaviour of the classifier indicate the applicability of fuzzy logic in the decision-making process, which takes place in an environment of inaccurate information and where the reliability of the results obtained by traditional fault diagnosis methods is inadequate.

Optimization of electric load clipping

Typical optimization methods include enumeration methods and methods based on dynamic programming. Many heuristical methods have also been reported. The enumeration and dynamic programming methods are optimal or nearly optimal but their major drawbacks are execution times. Heuristical greedy algorithms described in this work are fast when compared to those methods but not optimal. A dynamic successive programming method was developed to work with "time of use rates" (i.e. to work with hourly income and outcome rates for normal and peak loads, respectively). This method optimizes one load at a time. Time savings gained are 50-90% depending on how difficult the clipping situation is.

Neural networks in character recognition from x-ray images

Recognition consists of five phases, which are divided into the pre-processing stage and recognition stage. The pre-processing stage includes locating id-label, binarization, character segmentation and feature extraction. The recognition stage recognises characters from feature vectors. The binarization phase transforms the gray-level image into a black and white image. The segmentation phase locates individual characters from the id-label. The feature extraction phase calculates suitable features from character images. The recognition of feature vectors is performed with a multilayer perceptron neural network. Taking account the numerous variations and disturbances in x-ray images, the results are promising. The developed system was tested with 37 real images using resolutions of 120 dpi and 180 dpi. From 886 characters, 831 (94%) were correctly segmented and 79% of the segmented characters were correctly recognized.

Filtering ground reflection with neural networks

The effects of ground reflection were simulated by a theoretical model. Using the input-output data description generated by the theoretical model, multilayer perceptron neural networks were trained to respond to erroneous signals with correct elevation angles. The networks were then tested with an independent trajectory of a signal source. The theoretical results show that the disturbances can be filtered down to 0.4% from the original, when the number of the hidden layer neurons is around twenty. A disadvantage in the best simulations is the preprocessing of the signals that may be complex in practical applications. The results of the filtering simulations that did not demand the preprocessing of the signals were not as good.

3. Processing of video and audio signals (including hardware implementations)

Speech recognition

During 1996-1997 the goal of the research was to use neural networks and hidden Markov models for speaker-independent isolated and connected English digit recognition in noisy environments. The data set was spoken by 28 male speakers inside a car in three noise environments (motorway, city and a car park). In particular, the neural networks were researched as an alternative to the conventional conditional probability estimators such as mixtures of Gaussians. Hence, the purpose of the Multilayer perceptron was to estimate the emission probabilities for the conventional HMMs. Test set performance of the recognition system was 98.0% and 74.5%, respectively, in isolated and connected digit recognition at its best. The recognizer was tested with two training algorithms, embedded Viterbi (EV) and minimum classification error algorithm (MCE). The performances were slightly better for isolated digits when using MCE, while the digit strings were recognized a bit more accurately with EV training.

Speech coding: quantizations of LP coefficients

The tasks of 1997 placed emphasis on the implementation and testing of a generic VQ design platform. A resulting software product has been delivered to Nokia Research Center. The environment has been used for the estimation of a wide range of different LP (Linear Predictive) quantizers for an ACELP speech coder. Functionality of the quantizers has been tested in a floating point environment and observations of the performance and error resilience characteristics of different models have been collected.

lmplementation and development of WI-speech coder

The waveform interpolation speech coder utilizes the fact that speech is almost periodic during voiced sections. Using these periodic waveforms it is possible to form a surface. This surface can be modified and downsampled in an encoder and correspondingly interpolated in the decoder. Using this basic idea it is possible to achieve a relatively low total bit rate, for example 2.4 or 4.0 kbit/s.

At the time of writing the coder is almost complete; only some parameters require fine-tuning. The speech quality achieved so far has been fairly good. This project continued in the IMPRESS project, and final results were reported in late 1999.

Video compression based on visual reception

The maps of the perceptibility of the errors were quite well correlated with the error frames. However, it should be noted that the results were dependent on the parameter value mentioned in the abstract and the value can be varied according to the observers preferences. It has also been noticed that if the movement in the video sequence is very intense, the blocking effect can be seen around the new intra blocks (edges between original and predicted blocks are annoying). On the other hand, it must be noted that the algorithm showed that the erroneous blocks were very often situated next to the old intra blocks. This means that the algorithm was clearly working well, because normally the remaining errors are next to or very near to the old intra blocks.

Grid line removal from digitized x-ray images

The filter which has been developed has proved to give good results. It causes some blurring of the images, but not too serious. Blurring can probably be removed by using a restoration technique, for example by Wiener filtering. This project continued in the IMPRESS project, and final results were reported in the late 1998.

Image quality analysis

In this work, a system for subjective assessment of blocking artefacts was developed. The assessment was obtained by using fuzzy reasoning and it was based on quantitative features that are computed from the image with simple algorithms. The aim was to develop a system which is computationally feasible for real-time implementation.

Fuzzy control of video encoding

The test demonstrated that the fuzzy control strategy did not overcome the conventional rate control strategy. Image quality was not greatly improved. There are two main reasons for this: first, the number of bits to be shared is really small in the low bit rate applications using the hybrid DPCM/DCT coding, and secondly, the coding standard H.263 sets limits for the control of the coding parameters. However, the bit rate can be reduced in the less relevant areas while maintaining the subjective image quality. The bit rate can be reduced by 10-15 % with a simple image analysis at low bit rates.

Vector quantization index assignment

A genetic annealing algorithm was applied to the index assignment problem in vector quantization. The proposed combination of an annealing, selection and recombination mechanism is new and possesses an advantage with regard to implementation on a single processor machine because it gives satisfactory performance when using a reasonably small population. The assessment of the proposed method using a vector quantization of the line spectrum frequencies of a speech coder, the experiments with the vector quantizers trained on the Gaussian data as well as the test on the quadratic assignment problem benchmarks show the superiority of the proposed model.

Project information

Participants

• Signal Processing, Tampere University of Technology, 1995-98
• Measurement and Information Technology, Tampere University of Technology, 1995-98
• Electronics, Tampere University of Technology, 1995-96
• Machine Automation, VTT Automation, 1995-98
• Computational Engineering, Helsinki University of Technology, 1997-98

Project dates

Project volume

Project manager

Publications

Papers

2. A. Kantsila, M. Lehtokangas and J. Saarinen, "Adaptive Equalization of Binary Data Bursts Using Neural Networks", Proc. of 5th European Congress on Intelligent Techniques and Soft Computing, EUFIT'97, Aachen, Germany, Sep 8-11, 1997, Vol. 1, pp. 593-597.

3. A. Kantsila, M. Lehtokangas and J. Saarinen,"Adaptive Equalization of Binary Data Bursts", Proc. of IASTED International Conference on Signal and Image Processing, SIP'97, New Orleans, Louisiana, USA, Dec 4-6, 1997, pp. 117-122.

4. Kumpulainen Pekka, Imeläinen Heikki, and Joronen Tero, "Modelling Process with Data Modified Using Self Organizing Map," Nordic MATLAB Conference Final Program & Proceedings. Norra Latin, Conference Center Stockholm. Edited by Magnus Olsson. Computer Solutions Europe Ab, Stockholm. October 27-28, 1997, Stockholm, Sverige. pp. 11-109-112.

5. P. Kolinummi, T. Hämäläinen, H. Klapuri and K. Kaski, "Mapping of Multilayer Perceptron Networks to Partial Tree Shape Parallel Neurocomputer," Proceedings of the International Conference on Artificial Neural Networks (ICANN96), Bochum, Germany, July 1996.

6. Kolinummi, P., Hämäläinen, T. D. and Kaski, K., "Designing a Digital Neurocomputer; Improving ANN Processing with a Dedicated Hardware System," IEEE Circuits & Devices, The Optoelectronics Magazine 13, IEEE. 2, pp. 19-27.

7. Kolinummi, P., Hämäläinen, T. and Saarinen, J., "Mapping of Radial Basis Function Networks to Partial Tree Shape Parallel Neurocomputer," In: Gerstner, W., Germond, A., Hasler, M. & Nicoud, J.-D. (eds). Artificial Neural Networks - ICANN '97, 7th International Conference, Lausanne, Switzerland, October 8-10, 1997, Proceedings. pp. 1259-1264.

8. Kolinummi, P., Hämäläinen, T. and Kaski, K., "Mappings of SOM and LVQ on the Partial Tree Shape Neurocomputer," Proceedings of ICNN'97, 1997 International Conference on Neural Networks, 8-12 June, 1997, Houston, Texas, USA. 2. s. 904-909.

9. J. Lahnajärvi, M. Lehtokangas and J. Saarinen, "Filtering Ground Reflection with Neural Networks," Proceedings of 5th European Congress on Intelligent Techniques and Soft Computing, EUFIT'97, Aachen, Germany, Sep. 8-12, 1997.

10. J. Lahnajärvi, M. Lehtokangas, and J. Saarinen, "Neural Networks for Filtering Ground Reflection," The Yearbook of the Finnish Statistical Society 1996, 1997.

11. M. Lehtokangas, P. Salmela, J. Saarinen, and K. Kaski, "Weight Initialization Techniques in Neural Network Systems and Their Application," In C. Leondes (ed.), Algorithms and Architectures, vol. 1 in the Neural Network Systems Techniques and Applications Series, Academic Press, 1997.

12. V. Ruoppila, M. Vaalgamaa and M. Vuolahti, "On Subjective Quality in Quantization of LPC Parameters." submitted to IEEE Nordic Signal Processing Symposium (Norsig'98) 8-11 June 1998, Visgo Holiday Resort, Denmark.

13. P. Salmela, S. Kuusisto, J. Saarinen, K. Laurila and P. Haavisto, "Isolated Spoken Number Recognition with Hybrid of Self-Organizing Map and Multilayer Perceptron," Proceedings of International Conference on Neural Networks, ICNN'96, Washington, USA, Vol. 4, 1996, pp. 1912-1917.

14. P. Salmela, S. Kuusisto, J. Saarinen, K. Laurila and P. Haavisto, "The Hybrid of Self-Organizing Map and Multilayer Perceptron in Isolated Spoken Number Recognition," Proceedings of World Congress on Neural Networks, WCNN'96, San Diego, USA, 1996, pp. 63-68.

15. P. Salmela, M. Lehtokangas and J. Saarinen, "Improving Convergence of MLP with MC Initialization in Speech Recognition Application," Proceedings of the European Congress on Intelligent Techniques and Soft Computing, EUFIT'97, Vol. 1, 1997, pp. 490-494.

16. P. Salmela, M. Lehtokangas, Kari Laurila and J. Saarinen, "Discriminative Isolated Digit Recognition in Noisy Environments using Multilayer Perceptron Networks", Invited Special Session on Robust Speech Recognition, International Conference on Neural Information Processing, ICONIP'97, Vol. 2, 1997, pp. 1116-1119.

17. J. Seppälä, P. Korpisaari, V. Varjonen, M. Lehtokangas and J. Saarinen, "Recognition of X-ray Image's ID-label with Neural Network," The Yearbook of the Finnish Statistical Society 1996, 1997.

18. J. Seppälä, P. Korpisaari, V. Varjonen, M. Lehtokangas and J. Saarinen, "Character Segmentation from X-ray Image's ID-label," Proceedings of the 5th European Congress on Intelligent Techniques and Soft Computing, EUFIT'97, Aachen, Germany, Sep. 8-1 1, 1997.

19. Suojärvi Mika, Ritala Risto, Ihalainen Heimo, Jokinen Heikki, "Finding Significant Changes from Process Data", The yearbook of the Finnish statistical society 1996. Part I . Neural Networks and Statistics. Annual Conference of Statisticians, 7.-8.11.1996. Hakapaino Oy, Helsinki 1997. pp. 75-83. ISSN 0355-5941.

20. Suojärvi Mika: Neural network for estimation of quality. XIV IMEKO WORLD CONGRESS. New Measurements - Challenges and Visions. Vol. XA. Proceedings. Editor: Jouko Halttunen. June 1-6, 1997, Tampere, Finland. pp. 48-52. ISBN 952-5183-04-1.

21. Söderholm Kaj and Ihalainen Heimo, "Signal Feature Extraction Using Fuzzy Gates," Nordic MATLAB Conference '97. Final Program & Proceedings. Norra Latin, Conference Center Stockholm. Edited by Magnus Olsson. Computer Solutions Europe Ab, Stockholm. October 27-28, 1997, Stockholm, Sverige. pp. II-1-6.

22. Vehviläinen Pekko and Jokinen Heikki, "Determining the position of an edge of a dryer fabric by using neural and fuzzy computing," The yearbook of the Finnish statistical society 1996. Part I. Neural Networks and Statistics. Annual Conference of Statisticians, 7.-8.11.1996. Hakapaino Oy, Helsinki 1997. pp. 83-90. ISSN 0355-5941.

23. Vehviläinen Pekko: Älykäs paikan määrittäminen. Teema: Automaatiopäivät 97. Automaatioväylä, 5/1997. ISSN 0784 6428. Automaatioväylä Oy. pp. 20-22. (In Finnish)

24. Vehviläinen Pekko and Jokinen Heikki, "Processing the position indicator signal by neural and fuzzy computing," VTT SYMPOSIUM 172. Comadem'97. 10th International Congress and Exhibition on Condition Monitoring and Diagnostic Engineering Management. Vol. 2. Edited by Erkki Jantunen. June 9-11, 1997, Espoo, Finland. pp. 132-141. ISBN 951-38-4563-X, ISSN 0357-9387, UDC 658.58:061.3.

25. Vehviläinen Pekko, Jokinen Heikki, Saloniemi Jyrki, Vähätalo Harri: Determining the Position of a Dryer Fabric by Neuro-Fuzzy Computing. CD ROM, networking, (6 pp.), ISBN 952-5183-07-6, CD-ROM/esitelmä. (11) 3/4: Viiran paikan määrittäminen neuro-sumealla järjestelmällä. Automaatio 97/esitelmien lyhennelmät, (s. 49), ISBN SAS-21, Automaatio 97, Helsingin Messukeskus. Suomen Automaatioseura ry, Helsinki, Finland 23.-25.9.1997.

26. A. Vesterinen, A. Särelä, P. Vuorimaa, and M. Lehtokangas, "Anestesiakoneen sumea vikadiagnostiikka," in Suomen Tilastoseuran vuosikirja 1996, Helsinki, 1997, pp. 91-102.

27. M.Vuolahti, V. Ruoppila, M. Vaalgamaa, J. Saarinen, "Numerical study on predictive quantization of line spectrum frequencies." Proceedings of IASTED Conference SIP -97, Louisiana, USA.

28. M. Vuolahti, V. Ruoppila, "Performance of Predictive Vector Quantizers in Noisy Channels." Submitted to IEEE Nordic Signal Processing Symposium (Norsig'98) 8-11 June 1998, Visgo Holiday Resort, Denmark.

29. E. Yli-Rantala, T. Ojala, and P. Vuorimaa, "Vector quantization of residual images using self-organizing map," in Proc Int. Conf. Neural Networks, ICNN'96, Washington D.C., USA.

30. E. Yli-Rantala, T. Ojala, and P. Vuorimaa, "Vector quantization of residual images using self-organizing map with sample weighted training," in Proc. 4th European Congress on Intelligent Techniques and Soft Computing, EUFIT'96, Aachen, Germany, Sept. 2-5, 1996, pp. 325-328.

Laboratory reports

2. P. Kolinummi, T. Hämäläinen, J. Saarinen, K. Kaski, Kaupalliset Neurotietokoneet, Raportti 4-96, Tampereen teknillinen korkeakoulu, 1996. (In Finnish.)

3. M. Lehtokangas, P. Salmela, J. Saarinen and K. Kaski, "Weight Initialization Techniques in Neural Network Systems and Their Application," Research Reports B 12, Laboratory of Computational Engineering, Helsinki University of Technology, 1997.

4. P. Salmela, Neural Networks in Spoken Number Recognition, Report 1- 1996, Tampere University of Technology, Electronics Laboratory, 1996, 99 pp.

5. P. Salmela, Spoken Digit Recognition with Neural Networks and Hidden Markov Models, Report 3-1997, Tampere University of Technology, Signal Processing Laboratory, 1997, 140 pp.

6. Suojärvi Mika: Laatusuureen mallinnus neuroverkkojen avulla. Raportti TTKK/MIT, Tampere 1996, 54 s. ISBN 951-722-552-0. ISSN 1236-3227. (In Finnish.)

7. Suojärvi Mika: Signal features behind web break frequency. Report 39 (laboratory series), TTKK/MIT, Tampere 1996, 18 pp.

8. Vehviläinen Pekko: Determining the position of a dryer fabric by soft computing. Report 40 (laboratory series), TTKK/MIT, Tampere 1997, 56 pp.

Dr. Tech. Theses

2. M. Lehtokangas, Modelling with Layered Feedforward Neural Networks, Dr.Tech. Thesis, Tampere University of Technology, 1995, 228 pp.

3. P. Vuorimaa, Fuzzy Self-Organizing Map, and It's Applications, Dr.Tech. Thesis, Tampere Univeristy of Technology, 1995, 159 pp.

Lic.Tech. Theses

2. M. Vuolahti, Thesis for the Degree of Licentiate of Technology is in preparation.

M.Sc. Theses

2. Haukijärvi Mikko, "Fuzzy Control for Video Encoding in Real Time Applications," Tampere University of Technology, 1997.

3. J. Heikkinen, Sumea luokitus juotosliitosten vikadiagnostiikassa, diplomityö, Tampereen teknillinen korkeakoulu, 1996. (In Finnish.)

4. A. Kantsila, Adaptive Equalization of Binary Data Bursts, Master of Science thesis, Tampere University of Technology, 1997.

5. P. Kolinummi, Neurotietokoneiden arkkitehtuureja ja toteutusvaihtoehtoja, diplomityö, Tampereen teknillinen korkeakoulu, 1996. (In Finnish.)

6. H. Laine, "Spectroscopic Concentration Determination using Neural Networks," Master of Science thesis, Tampere University of Technology, 1995.

7. J. Lahnajärvi, Maaheijastuksen suodatus neuroverkolla, diplomityö, Tampereen teknillinen korkeakoulu, 1996. (In Finnish.)

8. M. Onnia, M.Sc. Thesis, in preparation.

9. J. Peltoniemi, Intelligent Methods for Chemical Agent Mixture Detection, Tampere University of Technology, 1996.

10. P. Salmela, Neural Networks in Spoken Number Recognition, Master of Science Thesis, Tampere University of Technology, 1996, 99 pp.

11. J. Seppälä, Tekstin tunnistus röntgenkuvan ID-kentästä neuroverkolla, diplomityö, Tampereen teknillinen korkeakoulu, 1997. (In Finnish.)

12. Suojärvi Mika, Laatuanturin signaalin käsittely neuroverkkojen avulla (Signal processing of a quality sensor with neural networks). TTKK, Sähkötekniikan osasto, 1996. (In Finnish.)

13. Söderholm Kaj: Tool for feature extraction using fuzzy gates. TTKK, Sähkötekniikan osasto, 1997.

14. M. Tammi, M.Sc. thesis, in preparation.

15. Vahteri Joni, "Fuzzy Assessment of Blocking Artifacts in MC/DCT Video Coding," Tampere University of Technology, 1997.

16. Aki Vehtari, "Neural network modeling of pumping process," M.Sc. Thesis, Laboratory of Computational Engineering, Helsinki University of Technology, 1997. (In Finnish.)

17. Vehviläinen Pekko, Determining the position of a dryer fabric by soft computing, TTKK, Automaatiotekniikan osasto, 1997.

18. Vesterinen Arto, Fuzzy Logic in Anaesthesia Machine Fault Diagnosis, Tampere University of Technology, 1997.

19. M. Vuolahti, WWW-palvelimen lukija-analyysi, diplomityö, Tampereen teknillinen korkeakoulu, 1996. (In Finnish.)

20. E. Yli-Rantala, Self-Organizing Maps in Codebook Design for Vector Quantization, Tampere University of Technology, 1996.

jukka.iivarinen@hut.fi
http://www.cis.hut.fi/neuronet/Tekes/3.shtml
Thursday, 30-Nov-2000 10:34:26 EET