Sharif Digital Repository

Keyword spotting with high speed and accuracy is an important subject whithin speech processing domain especially when we are dealing with various transmission channels. In this research discriminative keyword spotting methods are compared with HMM-based approaches. We have employed the discriminative approaches as our baseline methods due to their higher accuracy. The drawback of the conventional discriminative methods is their high computation cost and long execution time. The discriminative approach consists of two steps: feature extraction and classification. We have proposed four ideas to improve the performance of the baseline method. To improve the speed of the process, in feature... 

Transformer are one of the most important equipment in transmission and distribution network. Transformer unplanned outage have severe impacts on the continuity of power system operation and is also an irreparable economic harm to power network operators. To improve the reliability of transformers and to achieve an optimum operation cost, online condition monitoring is inevitbale. Information about the quality of the transformer insulation system is known as the best parameter to be monitored in transformer. Since partiale discharge (PD) signals are initiated long before the beginning of a severe damage, monitoring and its evaluation can be employed to warn the operator. Data mining on the... 

Memory and recalling process has always been a basic question. Decoding the Long-Term_Memory is one of the first steps in answering this question. Since various experiments in the field of human long-term memory, was conducted. This research is motivated by a trial that in which, the Mgntvansfalvgram (MEG) has been recorded while recalling the color and orientation of a grading which is associated with an object, after the object has been shown. High accuracy in Decoding the mentioned color and direction, will be decoding the long-term memory. In order to enhance memory decoding, the research studies different classifiers such as sparse based classifiers and other popular one. It has also... 

Target recognition techniques based on micro Doppler phenomenon are popular because they are applicable even on low resolution radars, in contrast to other techniques such as High Resolution Range Profile (HRRP) which need high resolution in range or angle. Usually, main purpose of such techniques is generating robust features against target initial state, velocity, aspect angle, etc. rather than features which exactly identify a target. Main approaches in the literature are based on time-frequency transforms (TFT) such as spectrogram in order to generate features to classify targets, but in this thesis, we propose a totally different method using Recurrence Plot for generating features... 

Digital video is one of the biggest part of digital data. The first step of digital video analytics is shot boundary detection. We used overlapped partitioning beside color histogram in uncompressed data and macroblock type prediction in compressed data as feature and supervised classifiers for decision making. Tests on TRECVID 2006 shows 8.9% improvement of F-measure in uncompressed video and 5.3% in h.264 bitstream. Supplementary test is done on IRIB dataset which shows 5.7% improvement of F-measure in uncompressed and 3.2% in H.264. H.264 based algorithm is almost 7 times faster in comparison to the algorithm that includes decoding  

Nowadays spam has become as a universal problem which all email users are familiar with it. Studies show that a large proportion of sent emails are spam. Obviously it results in wasting a vast range of resources. There is different ways to fight spam; each of them has its own strengths and weaknesses. The most common filtering technique is content based filtering. This problem has been addressed as a text classification problem. Two main defect of spam filtering techniques are manually definition of rules and circumventing them, one solution for overcoming this problem is applying machine learning algorithms. Spam classification using machine learning techniques is very successful and... 

Knowledge graphs are structured data sources which are widely used in the information process techniques. There are general and specialized knowledge graphs out there. These graphs will be used as the kernel of future search engines. Due to the lack of proper and tested Persian knowledge graphs, a method for knowledge graph extraction from news sources of the web has been introduced in this research.A knowledge graph extraction system from the unstructured web sources has been implemented in this research. In order to achieve this, a training dataset for the classifier was first extracted from semi-structured data of Wikipedia pages. At that time sentences were extracted from the... 

In this project we proposed a new method for improving the detection abality of a steganalyser with a pre-processing on contents of an image. Steganalysis, using machine learning, is designing a classifier with two classes: Stego or Cover. This classifier should be trained with extracted features from signal. The result of the training procedure is a machine that decides a signal belongs to stego or cover class. The first step of steganalysis process is extraction of proper features from signal. Proper feature is a variable that represents all of the useful properties of signal. Second step of this process is classifying data to two class of stego and cover. Many algorithms are proposed for... 

Transformers are the key component in power system transmission and distribution networks. Condition based maintenance will increase their expected life and online monitoring is essential to ensure operation reliability. In this work a new approach to transformer online monitoring is provided based on partial discharge (PD) measurement.Multi-source PD signal separated using time-frequency S transform (ST) that is applied to the PD signal waveforms. The resultant ST matrix is then converted to gray scale image from which high level features are extracted using Bag of Words (BoW). Gaussian mixture model (GMM) clustering is used to discover clusters in the feature space. For recognition of... 

Content-based spam filtering problem is defined as classifying input emails into spam and legitimate emails. so it is considered as an application of supervised-learning. The supervised learning methods often require a large training set of labelled emails to attain good accuracy and the users should label huge amount of emails. In reality, it is not reasonable to expect users to do this. To address this issue and reduce number of labelling request from user active learning techniques can be used. The goal of active Learning algorithms is to achieve appropriate accuracy by using fewer amounts of labelled data in comparison with supervised-learning methods.In this thesis two active learning... 

Detection degradation caused by distribution discrepancy between the training and testing domains is a common problem in object detection systems. The difference between training and testing domains’ distribution mainly happenes because of the different ways of collecting and gathering data. For instance, datasets which have images with different illumination, view point, resolution, background and are obtained by different acquisition systems, have variance in distribution. The solution toward improving the detection rate of the classifier trained on training (source) domain when it is applied on testing (target) domain is to use Domain Adaptation (DA) techniques. One of important branches... 

It is not possible to compute a linear classifier to classify real world images, which are the focus of this thesis. Therefore, the space of such images is considered as a complex. In such cases, kernel trick in which data samples are implicitly mapped to a higher dimension space, leads to a more accurate classifier in such spaces. In kernel learning methods, the best kernel is trained for the classification problem in hand. Multiple Kernel Learning is a framework which uses weighted sum of multiple kernels. This framework achieves good accuracy in image classification since it allows describing images via various features. In the image input space which is composed of different extracted... 

Nowadays, robustness of the Automatic Speech Recognition (ASR) systems against various noises is major challenge in these systems. Missing feature speech recognition approaches are our goal in this thesis for achieving robust ASR systems. In these approaches, low SNR regions of a spectrogram are considered to be “missing” or “unreliable” and are removed from the spectrogram. Noise compensation is carried out by either estimating the missing regions from the remaining regions in some manner prior to recognition, or by performing recognition directly on incomplete spectrograms. These techniques clearly require a "spectrographic mask" which accurately labels the reliable and unreliable regions... 

Recent researches show inefficient use of frequency spectrum such that there is shortage of frequency in operation. In order to overcome this problem cognitive radio are introduced that opportunist usage of frequency band is their prominent characteristic. Problems and challenges caused by using these networks are wide and increasing. In this thesis we focus on improving resource management algorithms in cognitive radio. In this study in order to have a dynamic and efficient management in choosing reliable and quality channels, reinforcement learning algorithms are used based upon existing data and experiences. Since this tool has the learning capability and analysis in dynamic situation of... 

The problem of stock price direction prediction is of great value among investors and researchers in the past decades. Even the smallest improvement in the performance of forecasting methods can lead to noticeable profit for investors. In this regard, in this research, a new method for filling the literature gap in the field of stock price direction forecasting is proposed. In the proposed method, two concepts of dynamics and model selection in dealing with data is investigated. Finally a predictive model is developed according to the two abovementioned concepts. Moreover, in this work, using a meta-learning approach one step towards making the prediction process automatic is taken. The... 

Concept drift is a common situation in data streams where distribution which data is generated from, changes over time due to various reasons like environmental changes. This phenomenon challenges classification process strongly. Recent studies on keeping a pool of classifiers each modeling one of the concepts, have achieved promising results. Storing used classifiers in a pool enables us to exploit prior knowledge of concepts in the future occurrence of them. Most of the methods presented so far, introduce a similarity measure between current and past concepts and select the closest stored concept as current one. These methods don’t consider possible relations and dependenies between... 

Detection and classification of arrhythmia is important especially for patients in Emergency care units. Early diagnosis of cardiac arrhythmia makes it possible to choose appropriate anti arrhythmic drugs, and is thus very important for improving arrhythmia therapy. Computer-Assisted Diagnostic (CAD) Systems are used in recent decades in which extracted features and classifiers are the most important factor. In this project, we try to focus on both of these two major factors in heart arrhythmia classification using HRV signal. Therefore, in this project, we try to classify different groups of arrhythmia using HRV signal processing especially the nonlinear processing. Our main aim is to... 

Power Quality can be simply defined as the quality of voltage at electrical loads. Detection and classification of voltage and current disturbances is of high importance in power system protection and monitoring. This procedure cannot be implemented by operators because of the high volume of the data which must be processed. So, it is needed to automate this procedure. Systems designed for this purpose usually contain three main parts: feature generation, feature selection and classifier design. The algorithms used for feature generation for power quality disturbances are mainly based on discrete Fourier transformation or discrete wavelet transformation. These approaches have shown some... 

Classification is one the contrivesial problems in machine vision and pattern recongnition. Traditional feature extraction methods which are based on low level feature extraction do not have high classification accuracy, thus they do not have the ability to represent images in feature space in discriminative way. In this thesis we have proposed a grid base method and used hidden Markov model (HMM) to include topological and spatial information in feature vectors. Then the classifiers created based on HMM feature extraction are combind. Combination of classifiers is based on designing a convex goal function. The goal of this optimization is to determine the wight of each classifier for... 

Classified ads are one of the first E-Commerce formed after the introduction of the internet to the world which is still popular. Divar is one of the main online classified ads and thousands of transactions are happening daily. The transaction of demandant and supplier happens in a risky relation that needs trust to facilitate this transaction. In this research, we are trying to study trust models in the context of C2C E-commerce. As there are no structural assurance and feedback mechanisms in Divar, we decide to study the initial trust of members due to a lack of information from each other. In this research, we gather data from 385 participants by an online questionnaire. The results show...