SIGNAL AND DATA PROCESSING
Year:2008 | Volume:- | Issue:1 (SERIAL 9)|Papers Count:7

The purpose of this study was to design and evaluate a target stimulus detector based on detecting P300 component. First, a suitable experiment was designed base on the oddball paradigm, so that a P300 wave is generated when subjects were confronted with Target stimulus. 20 subjects went through the designed paradigm and their respective brain signals were recorded. The best detection method was selected through the implementation and evaluation of some' proposed approaches on the recorded data. For the main processing block, which analyses the signal and makes a decision regarding the target or non-target stimuli for each subject, the proposed classifiers were LDA (linear Discriminant analysis) and decision tree. Also the optimal feature set was selected using a genetic algorithm method from a primary feature set including Mode frequency, Mean frequency, Median frequency, Discrete Wavelet transform coefficients and some Morphologic Parameters. Finally, the LDA was found as the best classifier. The final rate of correct detection of targets was 95% in the Loo (Leave One Out) method. Also the rates of correct classification of single trials were 71% for train data and 70% for test data. The best result was obtained using 18 selected features and the LDA classifier.

In this paper, using image processing and remote sensing techniques, rice yield estimation is performed for a region in Mazandaran province, North of Iran. Four Liss-III images from IRS-ID satellite have been used. Two of the images are from rice cultivation period and other two are from different times of the year. Geometric corrections have been performed on the images using ground control points. At first image pixels were classified to "rice" and "non-rice" classes using a2- class MLP neural network. Then rice farms were classified as "high yield" and "low yield" differentiations of rice cultivations using another MLP neural network. Estimation of rice yield for each of these classes of rice farms were then accomplished using an REF neural network model. Finally using the three mentioned neural network models, the estimation of rice yield for the whole region was performed Good results have been obtained from the models which are presented in the paper.

Syntax representation could be considered as the most important phase in developing an NLP system. Most of the grammars used in these systems are phrase structure ones that try to analyze the sentence by decomposing it into sub constituents. There are alternatives to this approach and link grammar as one of them is based on the words instead of constituents, hence called lexical grammar. We are going to assess their ability for Farsi. We do this by starting from the basic relations in Farsi and adding relations to them one by one to give a set of rules for trivial patterns. In each step, link grammar parser on an appropriate test suit tests the validity of the rules.

Affiixal structure of Farsi language allows a large amount of complex structures with various features that make the design of a complete lexicon rather impossible. But, according to have a fairly definite structure, this problem could be resolved by a morphological analyzer. In Farsi, adjacent morphemes rarely affect each other and on the other hand multiple affixes could appear on the same word, hence, Link grammar as a powerful formalism could be used for this purpose and makes the transformation from traditional to computational grammar more cost effective.Through introducing a system, which is able to decompose a word to its morphemes and reflects its features, we will give a complete survey on Farsi Morphology from a computational viewpoint. Unlike other approaches such as Finite State Transducers, the system covers derivational affixes as well as inflectional ones. Besides, we introduce an approach to represent features in this formalism. Knowledge of the system is listed, in details to make the approach most clarified.

Having been selected Rijndeal as the standard encryption algorithm by the National Institution of Standards and Technology (NIST) in October 2000, its hardware and software implementations have intensively been considered. In this paper, we propose two implementation models for Rijndeal based on Multibeat architecture on FPGA platform, which are efficient in hardware usage and throughput. In the first implementation, a Multibeat architecture based hardware unit is used to process the algorithm rounds by means of a feedback model. This implementation can be used in the medium rate applications. Further, this module can accept two data blocks simultaneously and use both parallel processing and resource sharing for processing of the inputs. The other implementation, designed based on Multibeat and pipeline model, is suitable for high rate applications. The analysis shows that both implementations have a suitable performances and the maximum throughput 38.272 Gbps can be achieved in the pipeline model.

Native speakers represent new concepts by semantic expansion or by creating new lexemes based on their language intuition. Usually unconscious word-formation is phonologically unmarked; therefore, there is a direct relation between the unmarkedness of the phonological structure of the lexemes and their frequencies. Based on the statistical analysis of 500000 Persian words, this study tries to give the spectrum of markedness in phonological structure and phonotactics of the lexemes. The findings of this study illustrate that in conscious word-formation if the newly formed word benefits an unmarked phonological structure, the native speakers will welcome it. Otherwise, conscious word-formation, at least phonologically, will not be successful.

It is well known that speech signal is affected by speaker's psychological stress. Some of the recent works have evaluated different acoustic features individually, for detecting stress in speech and among these parameters, the nonlinear feature of TEO-CB-Auto-Env is known as the best one. In this work, a new mixed feature (TEO-Pch-LFPC) is proposed and investigated for the task of stress classification, using simulated domain of SUSAS database (styles of Neutral, Angry, Loud and Lombard). The precedence of this work is that it uses more simple classifiers rather than HMM (i.e. static classifiers of KNN, LDA and SVM), and the Round Robin Method is exerted. For pair-wise classification, the proposed approach reaches 93.78% and in multi-style case, the accuracy of 70.22% is obtained.