مجله رایانش نرم و فناوری اطلاعات
سال چهارم شماره 2 (تابستان 1394)

Feature selection is one of the most important preprocessing steps for pattern recognition,data mining and machine learning. The objective of feature selection is selecting the most effectivefeature subset of original features set. So that in addition to dimension reduction of feature space,the recognition accuracy is improved. In this paper, feature selection process is the global search inthe space of all features using the wrapper approach. In this paper, the binary harmony searchalgorithm as intelligent algorithm is used to feature selection. Also, the one versus all method isused to classify and evaluate the selected feature subset.The performance of proposed approach is assessed on 6 datasets all of which are from UCI dataset.Numerous and Various experiments carried out on dataset reveal the superiority of the proposedmethod compared to similar approaches.

Network on Chip technology, NoC, is a perfect solution to control complex communication problem in system design. The first step in NoC-based designing for applications isselect NoC-topology such as mesh- topology. Operation mapping an application onto a mesh-based NoC is a NP-hard problem. So the mapping algorithms can be useful to reduce the algorithmic complexity as far as possible. In some mapping algorithms, priority selection for nodes is depend on select a neighbor node with high communication rate of pervious selected node. In these mapping algorithms, the edges with low communication rate are a problem. In this paper, the proposed mapping algorithm, tries to solve this problem with ignoring the low communication rate edges. To evaluate and compare the proposed algorithm, other mapping algorithms have been used such as Nmap, CastNet and Onyx. The proposed mapping algorithm has better result than other mapping algorithms for many applications such as VOPD and DVOPD.

Preventing voter coercion and vote buying are of the important challenges of any internet voting system. In proposed system, fingerprint image is used as a security password and the steganography technique is used to increase security of the registration phase. This proposed system make use of bank substructures for voting procedure and decrease the possibility of vote buying and prevent voter coercion by involving the voter’s banking data. These properties increase voters’ confidence to the voting system. The proposed protocol allows individual and universal verification and its counting method is expandable to the voting methods of different countries.

K-Nearest Neighbor (KNN) is one of the clustering methods. In this method, the distances between unknown sample with training samples are firstly calculated and the obtained distances are then sorted as ascending order and some of the smallest distances are selected. In this method, the K value is replaced by the number of the selections and after selection of the smallest K distances and the unknown sample is considered to a class which contains the highest number of the selected distances. The shortcoming of this method is that the error rate of clustering is become high due to noise and interference which can be removed by using the proposed method. In the proposed method, after sorting the distances as ascending order, the unknown sample is consideredto a class in which the K-minimum distance is related to that class for the first time. This is then followed by combination of the proposed method with Genetic Algorithm (GM) in training step andthen the resulted outputs are compared with the traditional methods. The obtaining results show that the proposed method results in lower classification error rate compared to traditional KNN and Genetic KNN (GKNN).

Using signal processing methods for fault detection of machinery is increasing nowadays. Noises added to the signal can negatively effect on efficiency of these methods. Time-domain averaging is a usual method to increase the strength of a signal. But, success of averaging is depends on an assumption which is corresponding points in averaging has same angle on the axle which is called synchronization. A little difference in synchronization can cause more decrease in efficiency of the method. Using tachometer is a usual way for synchronization. But, precision of tachometer isn’t enough. In this paper, as an attempt to solve this problem, a new method based on correlation is proposed for averaging which can synchronise data more precisely. Proposed method has been tested on real data gathered from a five-speed car gearbox. Using real-world data including a lot of components is an advantage of this research. Experimental results shows proposed method can reduce mean squared error from 0.419 to 0.103 in average which is a significant reduce.

Ensemble of classifiers is one of the famous methods to increase accuracy of Classifiers. Bagging, boosting and Error Correcting Output Codes (ECOC) are some of common methods to ensemble of classifiers. In this paper a new method for ensemble of Classifiers based on double combination is proposed and accuracy of this method in compare with other existing methods is evaluated for different datasets. Numerical results indicated that this method increases classifying accuracy in many cases comparing with other existing methods.

In today’s modern world, there is no place for passwords and security has the highest priority in current systems. Biometrics like face and voice can be circumvented by fraudulent methods but fingerprint has a high security in this aspect. In this paper a new fingerprint matching approach is introduced. The proposed method with various modifications has reached an accuracy equal or higher than the best existent methods. This method uses triangular relation of minutiaes. Experimental results show high accuracy and robustness to image transition and rotation. The time consuming part of the system are determined and then paralleled by CPU and GPU depend on dataor process complexity. GPU parallelization is performed by CUDA platform. Successful results of paralleled system show that the speed and accuracy of our system are suitable for real-time applications. We have managed to achieve up to 9x speed-up in the system run-time. The proposedsystem accuracy on FVC2002 database is higher than 99%.

Extracting the digital terrain model and automatic object detection using airborne and spaceborne data are still the most important issues in photogrammetry and remote sensing. In this research, an approach with three steps was presented in order to detect urban roads and buildings. In the first step, an initiative approach was proposed to extract the digital terrain model by utilizing the last range pulse of LiDAR. The second step contains determining some optimum features, gatheringtraining data from features, and employing a feed forward neural network with back propagation algorithm to recognize urban roads and buildings. In the third step, by carrying out a post processing on detection results, the accuracy of results would be improved. Moreover, in this research, the effect of Hough transformer for improving the buildings boundaries was analyzed. Finally, the detection results accuracy were evaluated by three criteria, including completeness, correctness and quality. Also, the results were compared with those of other methods that indicatedthe significant performance of the proposed approach of this study.

This paper presents a novel method for implementation of the elliptic curve scalar multiplication algorithm. The proposed method is based on sliding window method, recoding technique, parallel architecture, and a new high-radix scalable Montgomery modular multiplication. The sliding window method and recoding technique are utilized to reduce the number of required point operations. The parallel architecture is also utilized to perform the point addition and point doubling operations simultaneously. Moreover, the parallel architecture and a new high-radix scalable Montgomery modular multiplication in GF(2m) are utilized to speed up the multiplication. Our results show that the computation cost is reduced by about 83%-90% for w=4, and 93%-95% for w=8 in comparison with binary Karatsuba-Ofman, and 85% for w=4, 8 in comparison with parallel architecture in affine coordinate and projective coordinate.