فصلنامه پردازش علائم و داده ها
سال ششم شماره 1 (پیاپی 11، بهار و تابستان 1388)

This research has been carried out in the framework of Auto segmental-metrical (AM) phonology to study the stress in Persian. Two types of abstract and concrete prominences were distinguished in which the first one refers to the stress and the second one refers to the pitch accent. Stress is assumed to be a lexical property of the lexemes, but pitch accent is assumed to be an intonational element; therefore, contrary to stress, pitch accent in not fixed and predictable. Pitch accents could appear only on the lexically stressed syllables, if the context necessitates. Using the interface between phonology and morphology, it was concluded that all classes of lexemes are stressed on their final syllables in Persian; of course some grammatical morphemes are exceptions. Stress in inflected forms of words was formalized in which inflectional morphemes are stressed and receive the stress of the word, but clitics are unstressed and do not change the position of word stress if added to the stem. The results of this study could be used in linguistic studies and in different branches of Natural Language Processing.

Extracting semantic roles is one of the major steps in representing text meaning. It refers to finding the semantic relations between a predicate and syntactic constituents in a sentence. In this paper we present a semantic role labeling system for Persian, using memory-based learning model and standard features. Our proposed system implements a two-phase architecture to first identify the arguments by a shallow syntactic parser or chunker, and then to label them with appropriate semantic role, with respect to the predicate of the sentence. We show that good semantic parsing results, can be achieved with a small 1300-sentence training set. In order to extract features, we developed a shallow syntactic parser which divides the sentence into segments with certain syntactic units. The input data for both systems is drawn from RCISP corpus which is hand-labeled with required syntactic and semantic information. The results show an F-score of 81.6% on argument boundary detection task and an F-score of 87.4% on semantic role labeling task using Gold-standard parses. an overall system performance shows an F-score of 73.8% on complete semantic role labeling system i.e. boundary plus classification.

Current lie detection methods, based on the polygraph technique, rely upon the measurement of several physiological characteristics to discriminate whether a truth or a lie is expressed. P300-based GKT (guilty knowledge test) has been suggested as an alternative approach for conventional polygraph technique. The purpose of this study is to evaluate RT (response time) analysis and combination of RT and ERP to identify participants possessing specific guilty knowledge. For the analysis, several previous methods were implemented and results of RT analysis and fusion ERP and RT analysis were compared which other. The accuracy of RT analysis (bootstrapped analysis of reaction time method) is 81% and AUC (area under curve) of correct detection in guilty and innocent subjects is 0.85 which are comparable to 80% accuracy of ERP analysis (Wavelet classifier method) and AUC=0.86 in previous study. Between the many fusion methods, fusion of BART and BAD method has the better accuracy and best AUC, this is superior result compared to results of previous methods from McNemar’s test point of view. These results show that brain response and behavioral response have complement information. The Fusion of BART and BAD method is combination of brain and behavioral response, therefore this proposed as best approach for Assess "Guilty Knowledge".

Emotions play a powerful and significant role in human beings everyday life. They motivate us, impact our beliefs and decision making and would affect some cognitive processes like creativity, attention, and memory. Nowadays the use of emotion in computers is an increasingly in vogue field. In many ways emotions are one of the last and least explored frontiers of intuitive human-computer interactions. This can perhaps be explained by the fact that computers are traditionally viewed as logical and rational tools which is incompatible with the often irrational and seeming illogical nature of emotions. It is apparent that we as humans, in spite of having extremely good abilities at felling and expressing emotions, still cannot agree on how they should best be defined. until now, there are a bunch of good reasons which supports that emotion is a fitting topic for Human-Computer Interaction research. Human beings who are emotional creatures should theoretically be able to interact more effectively with computers which can account for these emotions. So Emotions assessed would make some improvement in HCI. The goal of our research is to perform a multimodal fusion between EEG’s and peripheral physiological signals for emotion detection. The input signals were electroencephalogram, galvanic skin resistance, blood pressure and respiration, which can reflect the influence of emotion on the central nervous system and autonomic nervous system respectively. The acquisition protocol is based on a subset of pictures which correspond to three specific areas of valance-arousal emotional space(positively excited, negatively excited, and calm). The features extracted from input signals, and to improve the results of brain signals, nonlinear features as correlation dimension, largest lyapunov exponent and fractal dimension is used. The performance of four classifiers: LDA, QDA, KNN, SVM has been evaluated on different feature sets: peripheral signals, EEG’s, and both. Synchronization likelihood is used as a channel selection algorithm and the performance of two feature selection algorithms; Genetic Algorithm and Mutual information is evaluated. The best result of accuracy in EEG signals is 63.3% with QDA as classifier, the best result of peripheral signals is 61.67% and the best of both is 63.3% with QDA. In comparison among the results of different feature sets, EEG signals seem to perform better than other physiological signals, and the results presented showed that EEG’s can be used to assess emotional states of a user. Also, fusion provides more robust results since some participants had better scores with peripheral signals than with EEG’s and vice-versa.

In this paper, we propose a comprehensive steganalysis scheme for JPEG images. In this method, the optimized features which can interpret high distinction between cover and stego images are extracted from images. These features have been selected after a careful study on modifications caused by different steganography algorithms on statistical characteristics of images. Furthermore, using a hierarchical decision making, we have considerably improved the detection accuracy. It has been shown that the first-order statistics of DCT coefficients (e.g. histogram) are often more successful than second-order statistics (e.g. different types of correlations) in detection of LSB flipping methods such as JSteg, OutGuess, JPHide&Seek and StegHide. On the other hand, the second order statistical characteristics have better performance in some other steganography methods (especially for LSB Matching, MB1, SSIS and PQ). The proposed Method reveals the weakness of the different steganography algorithms by thorough view on them. The results of our experiments indicate that the accuracy of proposed approach is better than some other state of the art steganalysis methods in the term of detection accuracy. Besides, it is more generalized and comprehensive. A database including 2000 JPEG images with different quality factors has been used for these experiments. The new scheme can detect six common steganography JSteg, OutGuess, F5, MB1, Sequential and Random LSB Matching, with accuracies higher than 80% for the payload of more than 20%. We have used SVM for our classification scheme.

P300 is known as the most prominent component between cognitive components in electrical brain activity. According to done research, when brain encounters an inconsistent stimulation during processing a series of usual stimulation, a P300 component appears in recorded brain signal which could distinguishes from usual ones. Amplitude of P300 decreases after a short during act of auditory simulation; so that we face difficulty in recognition of component features. In the research we considered reduction of the amplitude of P300 with five auditory stimulations and its reasons in three separate record blocks as well as recognition of the component with Neural Network and Genetic Algorithm. Finally single-trial recordings containing P300 component from single-trial recordings without P300 component have been discriminated by six optimum features as Neural Network classifier input in Pz channel with accuracy of 80.55% in learning data and 50% in test data in the first block.

In dark environments and foggy or smoky conditions where it is not possible to use eyesight and usual binoculars to detect human from other objects, the best solution is to use infrared images. This paper presents a robust method to recognize pedestrians in infrared image sequences. For this purpose, combination of SVM and histogram classifiers has been used. A pre-processing phase extracts image patterns similar to human patterns and delivers them to histogram and SVM classifiers. For training and testing phases of the presented algorithm thermal data base of OSU pedestrian video sequences has been utilized. Results of the algorithm present its good accuracy and performance.