The purpose of this research is to provide an efficient and effective way to classify ultrasound ‎imagery. These images are usually used to identify the type of cover and the ingredients of the earth. ‎Because the size of these images is large and they are produced in numerous frequency bands, they ‎are large in size and feature high-resolution vector. Also, the presence of noise, the difference in ‎sensor angle, light angle, atmospheric conditions, and many other factors caused the difference in ‎the spectral vector of the data belonging to a category, creating multiple sub-elements and, ‎consequently, non-linear features, so the classification of these pictures is challenging.‎‏ ‏‎ A Joint ‎Sparsity model for categorizing samples with nonlinear nature has a great deal of efficiency, as it ‎makes the spatial correlation between pixels exploitable. In this thesis, sparse representation based on ‎dictionary methos is used to determine the type of elements of each pixel. In the Joint Sparsity ‎model, the dictionary atoms are common to all pixels in each superpixel, while the coefficients are ‎different for each one, and the image is split into superpixels to extract the Joint Sparsity model. To ‎extract superpixels, spatial and spatial features simultaneously are used to select more homogeneous ‎regions, thus the spectrum of each pixel and the spatial resolution of the pixels are compared. ‎Superpixels were extracted using ASLIC and SSSE methods. Also, with the SOMP algorithm, the ‎sparse coefficients associated with each spectral vector are extracted and the dictionary atoms are ‎created by the K-SVD method. From the experiments and comparison with other approaches, it is ‎concluded that the superpixel segmentation as a pre-processing, and the use of spectral and spatial ‎properties in this segmentation, has a significant effect on the identification of homogeneous regions, ‎and thus the decomposition of a spectrum is done more accurately. On the other hand, labeling a ‎sample by an expert is a difficult and costly task for training a classifier. In this thesis, a semi-‎supervised method has been used to overcome the limited number of training samples. To do this, ‎the samples are used with two semi-supervised algorithms TSVM and S4VM, which take advantage ‎of different assumptions for categorization and the result is compared with each other. The results ‎showed that the proposed method achieved a total accuracy of 94.73% for the TSVM classification ‎and 97.13% for S4VM. This level of overall accuracy indicates that the proposed method's accuracy ‎has grown compared with supervised methods. It also suggests that semi-supervised methods with ‎limited training data have higher performance than supervised methods and can be generalized well ‎for test data if the appropriate and more distinct features are extracted.‎

Vehicle Ad-hoc Networks (VANET) are a new generation of mobile ad hoc networks (MANET) in which moving vehicles have the role of mobile nodes and because of high mobility vehicles, Network topology is constantly changing. quick changes in the topology of the network are considered as a big challenge for routing in these networks ; while the routing protocols must be robust and reliable. AODV Routing protocol is one of the known routing protocols in vehicular ad hoc networks that uses the criterion of the minimum number of hops in a route to select a route.This measure alone decreases the network performance in many scenarios and finally prevents more stable routs from being selected. In this thesis, to resolve Considered challenge, AODV algorithm has been improved using fuzzy logic. for this purpose, will be added five parameters speed, direction, Link expiration time, Reliability link and fuzzy cost in AODV Route Request packages and by using them, a probability is obtained as output to select the link on rout. The Criteria used to evaluating the performance of the proposed method consist of the Packet delivery rate, Packet loss rate, Normalized routing load and Average of end-to-end delay. The results of the proposed method showed an improvement in the assesment criteria compared to the AODV routing protocol and Another method based on fuzzy called PFQ-AODV. 

the emergence of digital information era and rapid development of the Internet makes the information change gradually from paper form to the electronic one. This issue make the users capable to search the news and books in an electronic way. Thus, the existance of systems in order for information retrieval seems necessary. In order to do that, the present study suggests a system for text classification by means of semi supervised fuzzy clustering with the weighted feature vector. In the suggested method after preprocessing phase, genetic algorithm and then tf-idf are used for dimensionality reduction and based on results high discriminting power features are chosen. Suggested clustering algorithm increases the clustering system effectiveness by utlization of some labled samples , specially when documents are highly similar to each other. A wighted matrix is considered in the domain of c-w-fcm which is calculated according to the variance of each feature and changes by clustering algorithm progress simultaneously . Therefore high variated features have more effects on the clustering process. With regard to clustering method , euclidean distance is used but such a distance dose not consider the equal values for dimensions. Finally, the proposed system effectiveness is examined on the reuters dataset and results showed success this method in comparison with fuzzy clustering algorithm and Weighted fuzzy clustering algorithm for clustering multi-criteria assessment is known.

Cloud computing is an Internet based computing environment, where storage and computing resources are assigned between users according to theirs needs dynamically and using virtualization technology. Virtualization is an underlying infrastructure of cloud computing, and it creates certain security problems during the development of cloud computing. One essential but formidable task in cloud computing is to detect malicious attacks and their types. Due to increasing incidents of cyber-attacks, design and implementation of effective intrusion detection systems to protect the security of information systems is essential. In this thesis, a host-based intrusion detection system for protecting virtual machines in the cloud environment Is proposed. To this end, first, important features of each class are selected using logistic regression and next, these values are improved using regularization. Then various attacks are classified using a combination of different classifiers: neural network, decision tree and linear discriminate analysis with bagging algorithm for each class. The proposed model has been trained and tested using the NSL-KDD dataset. For implementation of the model Cloudsim is used which, compared to other methods shows acceptable accuracy of % 97.51.

 Lung cancer is one of the hardest and most dangerous types of known cancer in the world. Patient survival is highly correlated with early detection. Such that the five year survival rate is increased. Some early types of lung cancer begin with a small mass of tissue within the lung, less than 3 cm in diameter, called a nodule. Most nodules found in a lung are benign, but a small population of them turn into malignant over time. Image Processing and Pattern Recognition techniques can have a significant role in detecting and diagnosing lung nodules.
In this thesis an automatic Computer Aided Diagnosis (CAD) system is proposed which contains two phases. In the first phase, lungs are separated well from the CT scan images according to the active contour segmentation method and next, based on the Sift features the proposed Bagging classifier, the lung images are classified to two classes of patient and healthy. In the second phase, according to a fully automatic Graph-Cut segmentation method the nodules are extracted from patient images and their diameters are measured. The size of a nodule is very effective in its classification to benign and malignant. Therefore, finally, nodules are classified to two classes of benign and malignant based on their size and texture features. The proposed methods are tested on CT scan lung images of LIDC database consortium. The performance of the proposed methods in detection and classification to benign and malignant classes according to reduced false positive and accuracy is comparable to previous solutions. Hence, this system could help physicians and radiologists as a second advice.

During the last decade, an increasing attention has been paid to communication in modern society. Todays, social networks with hundreds of millions of members are a powerful tool to guide the flow of information. Therefore, many researchers study on the various aspects of these networks. One of the important issues in social network analysis is community detection. This thesis uses the ant-based evolutionary algorithm which consists of three stages of exploration, construction and local optimization, fuzzy clustering approach is then applied to improve and fine-tuning the results. The difference between this study and previous methods is that the proposed algorithm can summarized in the use of ant-based algorithms, modularity as a criterion in evaluating the improvement of results at each iteration and fuzzy clustering.
To analyze the performance of the method presented in this thesis, two types of real-world and synthetic data sets were used. The results show that the community has been found by the proposed method is very competitive in comparison with the most prominent community detection methods in the literature on the issue of quality as achieved more accuracy between 0.2 to 0.5 percent in synthetic data sets and especially for some real-world networks the accuracy gained is about 1.98%. In general, on the LFR Benchmark our algorithm matches the previously best known values, in 71.9% of the instances; better than the best known values in 10.1% of the instances, and the second best values in 16.4% of the instances. For the remaining 1.6% of the instances the proposed algorithm finds the third best values.