Balázs Tusor

Human–Computer Interaction for Smart Environment Applications Using Fuzzy Hand Posture and Gesture Models

Ever since the assemblage of the first computer, efforts have been made to improve the way people could use machines. Recently, the usage of smart environments has become popular in order to make everyday living more comfortable and to improve the quality of living of humans. In this paper, a hand posture and gesture modeling and recognition system is introduced, which can be used as an interface to make possible communication with smart environment (intelligent space) by simple hand gestures. The system transforms preprocessed data of the detected hand into a fuzzy hand-posture feature model by using fuzzy neural networks and based on this model determines the actual hand posture applying fuzzy inference. Finally, from the sequence of detected hand postures, the system can recognize the hand gesture of the user.

Data Classification Based on Fuzzy-RBF Networks

Classification has been among the most typical computational problems in the last decades. In this paper, a new filtering network is proposed for data classification that is derived from radial base function networks (RBFNs), based on a simple observation about the nature of the classic RBFNs. According to that observation, the hidden layer of the network can be viewed as a fuzzy system, which compares the input data to the data stored in each neuron, computing the similarity between them. The output layer of the RBFN is modified in order to make it more effective in certain fuzzy decision-making systems. The training of the neurons is solved by a clustering step, for which a novel clustering method is proposed. Experimental results are also presented to show the efficiency of the approach.

A fuzzy hypermatrix-based skin color filtering method

In this paper, a classification method is proposed. The idea behind the classifier is the pre-calculation of fuzzy membership function values that are stored in fuzzy multidimensional arrays (so-called fuzzy hypermatrices) so their data can be accessed quickly run-time using the parameter values of the input data. This paper focuses on the 3 dimensional 2-class case in order to achieve rapid skin area identification based on pixel color using the images of a camera. A training algorithm is presented and the performance is illustrated by an image processing problem.

A fast fuzzy decision tree for color filtering

Fuzzy decision trees have been gaining popularity in the past two decades. They are the fuzzy extensions of crisp decision trees, introducing fuzzy logic into the nodes of the tree, thus making their generalization capabilities more robust. This paper presents a fuzzy decision tree architecture that is optimized for quick inference, in order to make the classification process as fast as possible. Furthermore, two training algorithms are presented to incrementally train fuzzy decision trees for realtime classification applications.

Circular fuzzy neural network based hand gesture and posture modeling

Recently, the usage of smart environments has become popular, in order to make everyday living more comfortable and to improve the quality of live of humans. In this paper, a hand posture and gesture modeling and recognition system is introduced, which can be used as an interface to make possible communication with the Intelligent Space by simple hand gestures. The system transforms preprocessed data of the detected hand into a fuzzy hand posture feature model by using fuzzy neural networks and based on this model determines the actual hand posture applying fuzzy inference. Finally, from the sequence of detected hand postures, the system can recognize the hand gesture of the user.

Active problem workspace reduction with a fast fuzzy classifier for real-time applications

In this paper, a Sequential Fuzzy Indexing Tables classifier is proposed for problems that require fast online operation. Its base idea comes from fuzzy hypermatrices (which are specialized versions of fuzzy look-up tables) that realize nearest-neighbor classification in order to recognize patterns similar to known ones. It is done by mapping the problem space into the memory in form of multidimensional matrices, so the class of the input data can be gained instantly in the evaluation phase. The downside of the base method is that the memory requirements scale exponentially with the number of attributes and the size of the attribute domains. The proposed classifier solves this issue for problems with large, but sparse workspaces by storing only a part of the problem domain. Thus instead of using a single multidimensional matrix, the classifier consists of a layered structure, breaking the multi-dimensional problem to a sequential combination of 2D fuzzy matrices.

A Fuzzy Shape Extraction Method

This chapter presents an easily implementable method of fuzzy shape extraction for shape recognition. The method uses Fuzzy Hypermatrix-based classifiers in order to find the potential location of the target objects based on their colors, then determines the areas where the most densely occurring positive findings in order to restrict the area of operation thus speeding the process up. In these areas the edges are detected, the edges are mapped to tree structures, which are trimmed down to simple outline sequences using heuristics from the Fuzzy Hypermatrix. Finally, fuzzy information is extracted from the outlines that can be used to classify the shape with a fuzzy inference machine.

Robust variable length data classification with extended sequential fuzzy indexing tables

Recurrent Neural Networks are widely used tools for the classification of variable length data. However, their training is generally a very time-consuming task, especially for problems with high dimensions. The classification method proposed in this paper aims to provide a fast and simple alternative. Extended Sequential Fuzzy Indexing Tables are following the principle behind lookup table classifiers in that they realize an input-output association by mapping the problem space using arrays. The proposed network achieves this by breaking the multi-dimensional problem space down to a sequence of combinations, resulting in a flexible architecture that can work well with varying length data.

A rule-based filter network for multiclass data classification

Nowadays, data classification is still one of the most popular fields of machine learning problems. This paper presents a new, adaptive, and easily applicable method for the solution of such problems. The method uses rules derived from the training data. The rules are processed by a rule-based inference network that is based on the classic Radial Base Function networks, with modifications in the output layer that change the functionality of the network. The training of the system, the appointing of rules is done by the clustering of the training data, for which two new clustering methods are presented and experimental results are shown in order to illustrate the efficiency of the system.

An input data set compression method for improving the training ability of neural networks

Artificial Neural Networks (ANNs) can learn complex functions from the input data and are relatively easy to implement in any application. On the other hand, a significant disadvantage of their usage is they usually high training time-need, which scales with the structural parameters of the networks and the quantity of input data. However, this can be done offline; the training has a non-negligible cost and further, can cause a delay in the operation. To increase the speed of the training of the ANNs used for classification, we have developed a new training procedure: instead of directly using the training data in the training phase, the data is first clustered and the ANNs are trained by using only the centers of the obtained clusters (which are basically the compressed versions of the original input data).