نشریه مهندسی برق و مهندسی کامپیوتر ایران
سال هجدهم شماره 1 (پیاپی 54، بهار 1399)

To communicate with people interactive systems often need to understand human activities in advance. However, recognizing activities in advance is a very challenging task, because people perform their activities in different ways, also, some activities are simple while others are complex and comprised of several smaller atomic sub-activities. In this paper, we use skeletons captured from low-cost depth RGB-D sensors as high-level descriptions of the human body. We propose a method capable of recognizing simple and complex human activities by formulating it as a structured prediction task using probabilistic graphical models (PGM). We test our method on three popular datasets: CAD-60, UT-Kinect, and Florence 3D. These datasets cover both simple and complex activities. Also, our method is sensitive to clustering methods that are used to determine the middle states, we evaluate test different clustering, methods.

AMP is a low-cost iterative algorithm for recovering signal in compressed sensing. When the sampling matrix has IID zero-mean Gaussian elements, the convergence of AMP is analytically guaranteed. But for other sampling matrices, especially ill-conditioned matrices, the recovery performance of AMP degrades and even may be diverged. This problem limits the use of AMP in some applications such as imaging. In this paper, a method is proposed for modifying the AMP algorithm based on Bayesian theory for non-IID matrices. Simulation results show better robustness properties of the proposed algorithm for non-IID matrices in comparison with previous works. In other words, the proposed method has more precision in recovery, and converges with less iterations.

Nowadays, the threats such as terrorism and cybercrime are extremely increased, therefore, the identity authentication process is very substantial for the national security of a country. In this paper, we propose a novel multimodal authentication system with sequential structure based on deep learning. In the proposed method, feature vectors are extracted automatically through deep network with an end to end architecture. A multi biometric system using two fingerprint and a face is implemented and evaluated. The results demonstrate that, the authentication is done by fingerprints in 91.42% cases and only for 8.58% cases the face modal is required. In addition, the proposed method is more accurate than first and second fingerprint by 35% and 30% at FMR=0.001, respectively. As a result, we augmented the accuracy of the system and at the same time reduced the acquisition and matching time. This conducts to the improvement of user convenience and security of the service provider, simultaneously. The achievements of this work can be used to increase the effectiveness of authentication process and can play an important role in the acceptability of real world applications.

This research is in the field of machine translation and in relation to extraction of Persian-English chunks from bilingual corpus by Spark. In this regard, the most important challenge is that the operation must be carried out on large corpus; therefore, it requires distributed computing along with big data analysis techniques and tools. When translating text, we are usually confronted with chunks that we need to find the corresponding chunks of each one in the target language and insert it in our translation; this is accomplished by locating it in a corpus that contain the chunks and their corresponding translations. The existing methods, perform this operations in a non-distributed way, therefore while they run slowly, they cannot use a very large corpus. To overcome this shortcoming, in this research a distributed method has been presented, which also takes distance between the sections of chunks into account. The proposed method extracts all possible chunks from the input sentences in the monolingual corpus and uses the correlation coefficient to translate those chunks using the bilingual corpus. We implemented the proposed algorithm in a platform consisting of a computing cluster with sixty-four GB of memory and a twenty-four-core processor in Spark. The incorporated experimental data was a Persian and an English monolingual corpus along with an English-Persian bilingual corpus, each of which containing 100,000 sentences. Experimental results show that run time could greatly be reduced, and the quality of translation is also significantly improved.

By increasing the precision of steganalysis attacks in discovering methods of steganography, the need to improve the security of steganographic methods is felt more than ever. The LSBM is one of the simplest methods of steganography, which have been proposed relatively successful attacks for its discovery. The main purpose of this paper is to provide a method for improving security of LSBM. The choice of the sequence of pixels to embed and how to modify them varies in LSBM-based methods. In most existing methods some of these decisions are made at random. In the proposed method in this paper, a multi-key idea in the first step and a genetic algorithm in the second step are used to make better decisions. In the proposed method, as MKGM, the image is blocked and GLSBM is executed for each block with different keys and finally the block with the least histogram change compared to the original block is included in the stego image. The GLSBM method is the same as the LSBM method except that the genetic algorithm is used to decide whether to increase or decrease non-matching pixels. Comparison of the image quality criteria and the accuracy of the attacks in the detection of the proposed method show that these criteria are improved compared to the original LSBM method.

With the social media engagement on the rise, the resulting data can be used as a rich resource for analyzing and understanding different phenomena around us. A sentiment analysis system employs these data to find the attitude of social media users towards certain entities in a given document. In this paper we propose a sentiment analysis method for Persian text using Convolutional Neural Network (CNN), a feedforward Artificial Neural Network, that categorize sentences into two and five classes (considering their intensity) by applying a layer of convolution over input data through different filters. We evaluated the method on three different datasets of Persian social media texts using Area under Curve metric. The final results show the advantage of using CNN over earlier attempts at developing traditional machine learning methods for Persian texts sentiment classification especially for short texts.

Heterogeneous networks have been regarded as an integral part of fifth generation communication networks in order to respond to the unprecedented growth of required data rates. In such networks, the existence of a variety of cells with base stations of varying capacities and transmit powers has enabled the repeated use of available bandwidth. Moreover, the excess load on the central base station can be directed to the sub-cell base stations. In the current work, a novel approach is proposed for such a load balancing problem in which some nodes previously connected to the main base station can be served by sub-cells through the use of some D2D relays. This will increase the overall network capacity, improve the quality of service (QoS) of cell edge users, and increase covered users. In this design, the maximization of the capacity of D2D links is formulated as an optimization problem which is not convex in general. To tackle this, the main problem is divided into two sub-problems of optimal resource allocation and user-relay pairing problems with much lower complexity. Simulation results demonstrate the superiority of the proposed method over existing works addressed in the literature.

This paper focuses on the effect of heterogeneous cache hierarchy in data center processors in the dark silicon era. For extreme-scale high performance computing systems, system-wide power consumption has been identified as one of the key constraints. As energy consumption becomes a key issue for operation and maintenance of cloud data centers, cloud computing providers are becoming significantly concerned. Emerging non-volatile memory technologies are favorable replacement for conventional memory. Here, we employ a nonvolatile memory called spin-transfer torque random access memory (STT-RAM) as an on-chip L2 cache to obtain lower energy compared to conventional L2 caches, like SRAM. High density, fast read access, near-zero leakage power and non-volatility make STT-RAM a significant technology for on-chip memories. In order to decrease memory energy consumption, it is required to address both the leakage and dynamic energy. Previous studies have mainly studied specific schemes based on common applications and do not provide a thorough analysis of emerging scale-out applications with multiple design options. Here, we discuss different outlooks consisting of performance and energy efficiency in cloud processors by running CloudSuite benchmarks as one of scale-out workloads. Experiment results on the CloudSuite benchmarks show that using STT-RAM memory compare to SRAM memory as last level cache, consumes less energy in L2 cache, around 59% at maximum.