فصلنامه پردازش علائم و داده ها
سال پانزدهم شماره 1 (پیاپی 35، بهار 1397)

Many efforts have been done to predict epileptic seizures so far. It seems that some kind of abnormal synchronization among brain areas is responsible for the seizure generation. This is because the synchronization-based algorithms have been the most important methods so far. However, the huge number of EEG channels, which is the main requirement of these methods, make them very difficult to use in practice. In this paper, in order to improve the prediction algorithm, the factor underlying the abnormal brain synchronization, i.e., the imbalance of excitation/inhibition neuronal activity, is taken into account. Accordingly, to extract these hidden excitatory/inhibitory parameters from depth-EEG signals, a realistic physiological model is used. The Output of this model (as a function of model parameters) imitate the depth-EEG signals. On the other hand, based on this model, one can estimate the model parameters behind every real depth-EEG signal, using an identification process. In order to be able to track the temporal variation of the parameter sequences, the model parameters, themselvese, are supposed to behave as a stochastic process. This stochastic process, described by a Hidden Markov Model formerly (HMM) and worked by the current researchists, is now modified to a State Space Model (SSM). The advantage of SSM is that it can be described by some differential equations. By adding these SSM equations to the differential equations producing depth-EEG signals, Kalman filter can be used to identify the parameter sequences underlying signals. Then, these extracted inhibition/excitation sequences can be applied in order to predict seizures. By using the four model parametetrs relevant to excitation/inhibition neuronal activity, extracted from just one channel of depth-EEG signals, the proposed method reached the 100% sensitivity, and 0.2 FP/h, which is very similar to the multi-channel algorithms. The algorithm can be done in an online manner.

This paper proceeds directions of arrival (DOA) estimation by a linear array. These years, some algorithms, e.g. Khatri-Rao approach, Nested array, Dynamic array have been proposed for estimating more DOAs than sensors. These algorithms can merely estimate uncorrelated sources. For Khatri-Rao approach, this is due to the fact that Khatri-Rao product discard the non-diagonal entries of the correlation matrix in opposed to Kronecker product. In this article, an algorithm named as Direction of Arrival (DOA) Estimation using Kronecker Subspace is proposed to solve more correlated sources than sensors via some properties of vectorization operator and Kronecker product. The simulations in different scenarios are presented considering various numbers of frames and correlation values, here. These verify our mathematical analysis. Furthermore, Cramer-Rao bound (CRB) which is a crucial criterion to estimate, is under investigating for DOA problem. Although, CRB for DOA estimation has been proposed before, it is applicable only for fewer sources than sensors. In this paper, CRB for more sources than sensor is derived by extending the dimensions with using both real and imaginary parts of the parameters. This bound is compared to the error of the presented algorithm. The simulations show that the error of the presented algorithm is merely 7 dB far from the CRB.

In fine-grained recognition, the main category of object is well known and the goal is to determine the subcategory or fine-grained category. Vehicle make and model recognition (VMMR) is a fine-grained classification problem. It includes several challenges like the large number of classes, substantial inner-class and small inter-class distance. VMMR can be utilized when license plate numbers cannot be identified or fake number plates are used. VMMR can also be used when specific models of vehicles are required to be automatically identified by cameras. Few methods have been proposed to cope with limited lighting conditions. A number of recent studies have shown that latent SVM trained on a large-scale dataset using data mining can achieve impressive results on several object classification tasks. In this paper, a novel method has been proposed for VMMR using a modified version of latent SVM. This method finds discriminative parts of each class of vehicles automatically and then learns a model for each class using features extracted from these parts and spatial relationship between them. The parts weights of each model are tuned using training dataset. Putting this individual models together, our proposed system can classify vehicles make and model. All training and testing steps of the proposed system are done automatically. For training and testing the performance of the system, a new dataset including more than 5000 vehicles of 28 different make and models has been collected. This dataset poses different kind of challenges, including variations in illumination and resolution. The experimental results performed on this dataset show the high accuracy of our system.

Cardiovascular diseases are the most dangerous diseases and one of the biggest causes of fatality all over the world. One of the most common cardiac arrhythmias which has been considered by physicians is premature ventricular contraction (PVC) arrhythmia. Detecting this type of arrhythmia due to its abundance of all ages, is particularly important. ECG signal recording is a non-invasive, popular method for an assessment of heart's function. Development of quick, accurate automatic ECG classification methods is essential for the clinical diagnosis of heart disease. This research analyzes the ECG signal to detect PVC arrhythmia. Different techniques are provided in order to detect this type of arrhythmia based on ECG signals. As these techniques use different methods for detection, the reaction of each one will be different to detect this type of arrhythmia. There is no classifier to give the best results for all matters at any time and combining classifiers improve the combined system results in comparison with each of the techniques.
In this study, the MIT-BIH arrhythmia database is used as a data source. Two datasets are used for training; the first contains 2400 samples, as in other studies, and the second contains 600 samples, including normal and PVC beats. Morphological features and features obtained from wavelet transform used in a combined classifier were used afterwards, which is the combination of the most common classifiers namely artificial neural network, SVM and KNN for PVC beat classification. Statistical significance features were selected using the p-value approach and normalized them. The best results were obtained when combining all three classifiers and using normalized statistical significance features. The designed hybrid system succeeded to detect PVC beats with 98.9±0.2% accuracy, 99.0±0.1% sensitivity, and 98.8±0.2% specificity. Also, the efficiency of the proposed method was shown when using limited training samples. The results showed the success of the proposed approach, specifically in comparison with other related research studies.

Awareness of other's opinions plays a crucial role in the decision making process performed by simple customers to top-level executives of manufacturing companies and various organizations. Today, with the advent of Web 2.0 and the expansion of social networks, a vast number of texts related to people's opinions have been created. However, exploring the enormous amount of documents, various opinion sources and opposing opinions about an entity have made the process of extracting and analyzing opinions very difficult. Hence, there is a need for methods to explore and summarize the existing opinions. Accordingly, there has recently been a new trend in natural language processing science called "opinion mining". The main purpose of opinion mining is to extract and detect people’s positive or negative sentiments (sense of satisfaction) from text reviews. The absence of a comprehensive Persian sentiment lexicon is one of the main challenges of opinion mining in Persian.
In this paper, a new methodology for developing Persian Sentiment WordNet (HesNegar) is presented using various Persian and English resources. A corpus of Persian reviews developed for opinion mining studies are introduced. To develop HesNegar, a comprehensive Persian WordNet (FerdowsNet), with high recall and proper precision (based on Princeton WordNet), was first created. Then, the polarity of each synset in English SentiWordNet is mapped to the corresponding words in HesNegar. In the conducted tests, it was found that HesNegar has a precision score of 0.86 a recall score of 0.75 and it can be used as a comprehensive Persian SentiWordNet. The findings and developments made in this study could prove useful in the advancement of opinion mining research in Persian and other similar languages, such as Urdu and Arabic.

Question answering system is a field in natural language processing and information retrieval noticed by researchers in these decades. Due to a growing interest in this field of research, the need to have appropriate data sources is perceived. Most researches about developing question answering corpus area have been done in English so far, but in other languages as Persian, the lack of these corpora is perceived. In this article, the development of a Persian question answering corpus called Rasayel&massayel will be discussed. This corpus consists of 2,118 non-factoid and 2,051 factoid questions that for each question, question text, question type, question difficulty from questioner and responder’s perspective, expected answer type in coarse-grained and fine-grained level, exact answer, and page and paraghraph number of answer are annotated. The prposed corpus can be applied to learn components of question answering system, including question classification, information retrieval, and answer extraction. This corpus is freely available for the academic purpose as well. In the following, a question answering system is presented on the Rasayel&massayel corpus. Our experimental result represents that the intended proposed system has achieved 82.29 % accuracy and 56.73 % mean reciprocal rank. It could be also claimed that this is the first ever question answering system and corpus with such features in Persian.

Blood pressure is one of the vital signs. Specially, it is crucial for some cases such as hypertension patients and it should be monitored continuously in ICU/CCU. It must be noted that current systems to measure blood pressure, often require trained operators. As an example, in post-hospital cares, blood pressure control is difficult except with the presence of a nurse or use of a device that minimizes the patient's involvement in the measurements. In this way, Photoplotysmography (PPG), which is a noninvasive method for pulse wave recording, seems to be ideal to make simple tools for blood pressure measurement in home care. In other words, it is so helpful or rather necessary to design a non-invasive, cuff-less, subject-independent system for blood pressure measurement.
In this study, two optical sensors were located on the finger and the wrist. Twenty healthy volunteers in different situations were examined to record PPG signals. Also, blood pressure values were measured by cuff-based noninvasive blood pressure system on left arm as a reference value. Recorded signals were filtered and processed in MATLAB R2014a software. To promote the estimation accuracy and subject-independency, 16 temporal features in addition to the pulse transit time (PTT) were extracted from the wrist PPG signal. To estimate blood pressure values, three neural networks were used as the estimator: Feedforward Neural Network (FFN), Redial Basis Function Neural Network (RBFN) and General Regression Neural Network (GRNN). After comparison of their results; the General Regression Neural Network was used for blood pressure estimation. The MSE errors estimated by the best estimator, were 0.11±1.18 mmHg and 0.15±2.3 mmHg for systole and diastole pressure respectively.

Phrase-boundary model for statistical machine translation labels the rules with classes of boundary words on the target side phrases of training corpus. In this paper, we extend the phrase-boundary model using shallow syntactic labels including POS tags and chunk labels. With the priority of chunk labels, the proposed model names non-terminals with shallow syntactic labels on the boundaries of the target side phrases. In comparison to the base phrase-boundary model, our variant uses phrase labels in addition to word classes. In other words, if there is no chunk label in one boundary, the labeler uses the word POS tag. The boundary labels are concatenated where there is no label for the whole target span. Using chunks as phrase labels, the proposed model generalizes the rules to decrease the model sparseness. The sparseness has more importance in the language pairs with a lot of differences in the word order because they have less number of aligned phrase pairs for extraction of rules. Compared with Syntax Augmented Machine Translation (SAMT) that labels rules with the syntax trees of the target side sentences, the proposed model does not need deep syntactic parsing. Thus, it is applicable even for low-resource languages having no syntactic parser. Some translation experiments are performed from Persian and German to English as the source and target languages with different word orders. In the experiments, our model achieved improvements of about 0.5 point of BLEU over a variant of SAMT.

Nowadays, digital signal processors (DSPs) are appropriate choices for real-time image and video processing in embedded multimedia applications not only due to their superior signal processing performance, but also of the high levels of integration and very low-power consumption. Filtering which consists of multiple addition and multiplication operations, is one of the most fundamental operations of DSPs. Therefore, there is a need for an additional unit just after the multiplication unit in DSPs. By combining multiply and add units, new structure named MAC (Multiply and ACcumulate) unit is provided. Residue Number System (RNS) can improve speed and power consumption of arithmetic circuits as it offers parallel arithmetic operations on each moduli and confines carry propagation to each moduli. In order to improve the efficiency of the MAC unit, RNS could be utilized.
RNS divides large numbers to smaller numbers, called residues, according to a moduli set and enables performing arithmetic operations on each moduli independently. The moduli set {2n-1,2n,2n} is the most famous among others because of its simple and efficient implementation. Among this moduli set, modulo 2n circuits are the critical path due to (n)-bit wide data path despite other two modules which all have n-bit wide operands. In order to overcome the problem of (n) bits operands, three representations has been suggested: diminished-1, Signed-LSB and Stored-Unibit. Although different multipliers have been proposed for diminished-1 representation, no multiplication structure has been proposed for the last two ones. Modulo 2n multipliers are divided into 3 categories depending on their inputs and outputs types: both operands use standard (weighted) representation; one input uses standard representation, while the other one utilizes diminished-1 representation; both inputs use diminished-1 representation. Although several multiply and add units have been proposed for the first 2 categories, no MAC unit is proposed for the multipliers of a third category which outperform multipliers of other categories. In this article at first, one modulo 2n MAC unit for the third category is proposed and then for further improvement, pipeline and multi-voltage techniques are utilized. Pipeline structure enables a trade-off between power consumption and delay. Whenever high-performance with least delay is desirable, nominal supply voltage can be chosen (high performance mode) otherwise by reducing supply voltage to the amount at which pipeline circuit and normal circuit without pipeline would have the same performance, power consumption decreases significantly (low power mode).
Simulations are performed in two phases. At first phase, proposed MAC unit without pipeline structure is described via VHDL code and synthesized with synopsys design vision tool. Results indicate that the proposed structure outperforms PDP (Power-Delay-Product) up to 39% compared to the state of the art MAC units. At second phase, CMOS transistor level implementation in two modes i.e. low power and high performance modes with Cadence Design Systems tool is provided. Simulation results indicate that at low power condition, proposed pipeline MAC unit yields to 71% power savings compared to existing circuits without declining efficiency. Furthermore, at high performance condition, however power consumption has increased, reducing delay up to 54% yields to 39% PDP savings for proposed pipeline MAC unit.

Deductive Database systems are designed based on a logical data model. Data (as opposed to Relational Databases Management System (RDBMS) in which data stored in tables) are saved as facts in a Deductive Database system. Datalog Educational System (DES) is a Deductive Database system that Datalog mode is the default mode in this system. It can extract data to use outer joins with three query languages (Datalog, SQL and RA) in default mode. In 2004, system DES was designed and implemented by Fernando S´aenz-P´erez from Department of Artificial Intelligence and Software Engineering, Complutense University, Madrid, Spain. In a paper, this researcher introduced outer joins of system DES in 2012. The most important objective of present research is to complement and extend the paper authored by mentioned researcher. Therefore, in prior research, choosing the most appropriate query language has not been investigated to use outer joins for data extraction in Datalog mode in DES system. In this study, by considering two parameters (cost of writing a query and memory usage of a query) choosing the most appropriate query language has been investigated to use outer joins for data extraction in Datalog mode in Deductive system DES. Cost of writing a query parameter is considered in this study to decrease the query typing time, but other parameters are related to the query processing are not considered. If the processing time of the three query languages is assumed identical, after entering the query in the system DES, the idea of the present study (reduction of the typing time) can lead to the reduction of the response time. Also, there are two hypotheses in this study as follows: 1) it is assumed that the user is fluent in all three query languages and wants to access the given data quickly through the most appropriate query language. 2) In the present study, the simplicity or difficulty of a query language is not considered. The results of the research show that one language cannot be appropriate for all queries; therefore, for every different query the most appropriate query language must be chose to use outer joints. In the current research, the most appropriate query language is the one in which, in comparison with other two query languages, the user will need to use less buttons of the keyboard to press in order to fulfill the query. The decrease in the number of buttons pressed by the user will decrease the time consumed to fulfill the query and, therefore, it will lead to a faster access to data.