Sharif Digital Repository

To fully understanding a text written in a natural language, we need to comprehend the events within that text. Temporal relation extraction always have been one of the main challenges in natural language processing in semantic level. Temporal relation extraction makes the understanding and interpretation of text easier and the extracted information can be used in many natural language systems like question answering, summarization, and information retrieval systems. Early researches on temporal relation extraction was mainly on English and limited to rule based systems. However, with extending the English corpora and availability of temporal corpora in other languages, more attention has... 

Fine-grained image classification is image classification where the considered classes are all sub-classes of a certain, more general class. In this setting of the problem, the classes are visually very similar to each other, such that an unskilled human cannot discriminate between them. In this case, proposed methods for the ordinary image classification problem do not obtain good classification accuracy. So proposing new methods for solving this problem is necessary. In this thesis two new methods, based on recent advances in deep learning are proposed for solving the fine-grained image classification problem. First by improving several parts of one of the recent proposed methods for this... 

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... 

Expert retrieval when the number of experts are limited is an open problem. Undoubtedly, becoming an expert in a field is a time consuming and expensive task; thus finding the best candidates is a crucial task. In addition, passage of time and growth of knowledge could change the view of a person towards life and his work, which may lead to the change of his or her field of work. When considering the changes each person makes in his or her life, it becomes obvious that they are not far from the original status. Therefore, recommending all possible options around a person could really help the task of decision making. This research is addressing two similar issues of finding experts, in a... 

Simultaneous development of technology and social networks and universal access to them caused to produce and distribute huge volume of videos that recognition of their content without use of machine vision is very hard. This thesis examine some video classification algorithms to improve them. The algorithm that is used to improve is based on one of local descriptor algorithms. At first with using STIP tools, the local interest point found by Harris3d and describe by HOG/HOF. Then by using Bag of Features, all local descriptors in a video produce a descriptor per videos. Bag of Features divide the domain of all local descriptors from all videos to K cluster and produce a vector per video... 

This thesis investigates the expertise ranking problem. Recently, the expertise ranking problem has attracted lots of attention in Information Retrieval community. The broad usages of expert ranking algorithms in commercial search engines indicate its importance and usability. Expertise ranking problem is concerned with finding people who are knowledgeable in a given topic. The main research questions in this thesis are related to three important questions related to expert ranking problem. The first question is what the sources of evidences are and how we can infer expertise of a person on a given topic. The second question is concerned with the modeling of information related to each... 

Distance Metric Learning algorithms have been widely used in Machine Learning methods recently. In these algorithms a distance function between objecs (data points) is learned based on their labels or similarity and dissimilarity constraints. Recent works have shown that a good precision is obtained in classification or clustering methods which use these functions. Since in the current systems many of data points do not exist at the beginning and are added to the training set as the algorithm is run, online methods are needed to update learned metric due to new data.
In this thesis, we proposed a new online distance metric learning method that has higher performance than existing... 

Object recognition is a field of image processing with the purpose of object detection and determining its pose. There was a wide research of object recognition using 2d information. These methods are less efficient due to their sensitivity to view and scale. Today with development of 3d technology, 3d reconstruction of objects has been eased, thus there is a new trend of 3d methods. 3d object recognition uses 3d data and model to recognize objects more efficiently in different views and in occlusion and clutter. The related works are classified in three category of view-based, feature matching, and classification. In this research we attempted to improve one of the best feature matching... 

The emergence of applications such as spam detection and online advertising coupled with the dramatic growth of user-generated content has attracted more and more attention to stream classification. The data stream in such applications is large or even unbounded; moreover, the system is often required to respond in an online manner. Furthermore, one of the main challenges of stream classification is that often the process that generates the data is non-stationary. This phenomenon, known as concept drift, poses different challenges to the classification problem.Therefore, an adaptive approach is required that can manage concept drift in an online fashion. This thesis presents a probabilistic... 

Nowadays, by growing computer networks and also increase in the power of processors we are faced with large scale data. Due to this fact, a branch of machine learning called data stream processing is assigned to investigate an appropriate solution for this problem. Classification is one of the issues in the data streams. In data stream classification problem we are faced with some challenges, so many researches have been done to solve some of them like infinite length of stream and concept drift. However most of the existing methods don’t consider novel classes which are one of the most important issues in processing of data streams. According to researches conducted in recent years, this... 

Recent advances in storage and processing have provided the ability of automatic gathering of information which in turn leads to fast and contineous flow of data. The data which are produced and stored in this way are named data streams. It has many applications such as processing financial transactions, the recorded data of various sensors or the collected data by web sevices. Data streams are produced with high speed, large size and much dynamism and have some unique properties which make them applicable in precise modeling of many real data mining applications. The main challenge of data streams is the occurrence of concept drift which can be in four types: sudden, gradual, incremental or... 

Image classification is one of the most challenging issues in computer vision. One sort of such classifications is Scene Classification. To perform automatic classification reserchers used many aproches.The general approach used features directly extracted from the image, such as color and texture or features extracted by the SIFT algorithmetc. Another method is based on recognizing object of the Scene (espessially indoor scene). This method is based on finding of a limited number of prespecified objects. In the proposed method, first a window surrounding each objects, (regardless of the type of object) founded. Then the SIFT feature is extracted from that window. All features (corresponding... 

Scene classification and understanding is one of the most important fields in computer vision. Its applications are such as exploring robot navigation enviroment, content-based image retrieval (CBIR), organization in image databases, highly semantic describing images and videos and content extraction of videos.Many methods and algorithm are proposed till today to deal with diversity of this field by emphesizing on feature based methods or machine learning based methods. In this research we have focoused on proposing a new algorithm which is using principals of NBNN image classification method but major changes in how to exract distance metric from Nearest neighbour and how to use local... 

The ability of measuring colors of objects, independent of light source illumination, is called color constancy which is an important field in machine vision and image processing. In this thesis, we propose five new combinantional ways in color constansy fields. The first two proposed methods use neural networks to combining basic methods. The third and forth proposed methods use fuzzy measures and integrals for combining color constancy methods. And finally fifth method combines methods with indoor outdoor classification this method has the best result (3.01 median angular error) in proposed methods and past methods in color constancy fields. It is shown in this article that the proposed... 

Nowadays, living in information era has forced us to face with a great deal of problems of which the input data is received like a nonstop endless stream. Intrusion detection in networks or filtering spam emails out of legal ones are instances of such problems. In such areas, traditional classification algorithms show function improperly, thus it is necessary to make use of novel algorithms that can tackle these problems. Among classification algorithms, decision trees have significant advantages such as being independent of any parameter and acting robust against outliers or unrelated attributes. Moreover, results of a decision tree are quite easy to interpret and extract rules from.... 

Image classification, i.e. the task of assigning an image to a class chosen from a predefined set of classes, has addressed in this thesis. At first the classifier is divided into two major sub partitions, feature extraction and classifier. Then we show that by using local feature extraction techniques such as BOW the classification accuracy will improve. In addition, using unlabeled data is argued as the fact to deal with high nonlinear structure of features. Recently, many SSL methods have been developed based on the manifold assumption in a batch mode. However, when data arrive sequentially and in large quantities, both computation and storage limitations become a bottleneck. So in large... 

Scene classification is one of the most controversial fields in computer vision. It has many applications such as robot navigation and control, content-based image retrieval (CBIR), semantic organization of image databases, depth estimation and multimedia services. In fact the outcome of any classification system depends on the ability of the feature vector defined for the problem, by means of its distinguishing strength. In this research we focus on efficient feature extraction methods. In recent years, methods based on bags of features and special pyramid approach, have shown good performance in scene classification comparison to the others. So we based our proposed method on these ideas.... 

In this thesis, we have discussed image classification by sparse representation. Sparse representation is used in two different ways for image classification. The first goal of sparse representation is to make an efficient classifier, that can learn the subspace, in which the data lies. In this field we have surveyed various methods. We also proposed a method, called ”Locality Preserving Dictionary Learning” that works approximately better than state of the art similar methods, specially when training data is limited. We have reported the result of lassification on four datasets including MNIST, USPS, COIL2 and ISOLET. Another use of sparse representation, is to extract local features from... 

Classification of hyperspectral remote sensing images is a challenging problem, because of the small number of labeled pixels, high dimensionality of the data and large number of pixels. In this context, semisupervised learning can improve the classification accuracy by extracting information form the distribution of all the labeled and unlabeled data. Among semi-supervised methods, manifold-based algorithms have been frequently used in recent years. In most of the previous works, manifolds are constructed according to spectral representation of data, while spatial dependency of pixel labels is an important property of the images in remote sensing applications. In this thesis, after... 

Event extraction is one of the important tasks in Natural Language Processing (NLP). Many NLP applications like question answering, information extraction and summarization need to have some knowledge about events of input documents. There are several definitions for events in NLP domains. In this dissertation, the event is viewed as an element in a network of temporal information. Therefore, the project has been based on the ISO-TimeML specification language, which is the standard scheme for temporal information processing in natural texts. Event extraction based on ISO-TimeML has been performed for a number of languages including English, French, Spanish, and Korean. However, for Persian...