Search

Expanding ubiquity of high-resolution digital video over the Internet calls for better compression methods to enable streaming with higher compression efficiency and lower latency. Recently, important gains have been achieved in learned image compression by using end-to-end learned models. However, these improvements haven't been fully leveraged in video compression. This paper aims to improve upon work proposed by Lu et al. in CVPR 2019, which has been claimed to outperform conventional video codecs in terms of PSNR and provide some implementation details that are absent in the original paper. Ultimately, we show that modeling latent symbols by Laplacian distribution outperforms the Gaussian assumption used in the original work and also demonstrate in a repeatable fashion that our learned model is superior to x264 video codec in terms of PSNR over a range of compression rates measured by bit-per-pixel.

Metric learning is a frequently utilized method that can be applied to many different computer vision problems and; thus, facilitates these problems. In this paper, the proposed method for manipulating the feature space with the help of quadruplets, which is one of the metric learning methods, is explained with details, and the results obtained by using it are shown. Many methods used in the field of metric learning, have been developed on the Siamese and triplet structures and satisfactory results have been obtained by them. Similarly, the quadruplet structure is still being studied and different approaches are proposed in the literature. How to select four elements for generating a quadruplet is not standardized at present. However, by looking at the mining methods of Siamese and triplet samples, it can be concluded that each sample in the dataset that is trained with quadruplet samples, must have at least two hierarchical labels. This makes it possible to train the quadruplet structure only on certain datasets that have the hierarchical labels. In this paper, an unsupervised method is proposed to use the quadruplet structure in datasets that do not have a hierarchical label structure. Besides, the performances of three different quadruplet mining methods are compared and an ablation study is conducted by discussing the advantages and disadvantages of the proposed method.

Millimeter wave (MMW) communication was proposed as a high-bandwidth wireless access solution to cope with the increasing user demand for mobile data and has been already standardized in 5G networks. As another highcapacity wireless solution with operation in unregulated optical spectrum, visible light communication, also known as LiFi, was further proposed in the literature. Hybrid wireless networks are particularly attractive to exploit the complementary features of LiFi and MMW technologies. In this paper, we consider a hybrid LiFi/MMW wireless communication system and investigate how to implement efficient vertical handover for improved network performance. Specifically, we propose an adaptive handover technique where handover parameters are determined based on network load and current data rates of users. Our results demonstrate that the proposed handover technique ensures a more balanced distribution of users between two underlying access technologies and improves overall network performance.

Optical wireless communication technology is one of the most important alternative schemes for radio frequency communication. Flip orthogonal frequency division multiplexing OFDM (Flip-OFDM) is a type of indoor optical wireless communication that transmits the positive and the absolute values of negative parts of the signal in two successive OFDM symbols. In this paper, the Fast Walsh Hadamard Transform (FWHT) based Flip-OFDM method (FWHT Flip-OFDM) is proposed to improve the performance of the Flip-OFDM system. Computer simulation studies are performed to test the performance of the proposed method over the bit error rate (BER) and peak to average power ratio (PAPR) performance criteria. From the obtained simulation results, it is observed that the proposed method is better than both VLC-OFDM and Flip-OFDM methods for both performance metrics.

Nowadays, multispectral images are acquired by many different satellites. These satellites produce a large amount of data by traveling around the world. Efficient compression of these images are important issues. In order to solve this problem, the most used method is compression of the images with JPEG2000 method. However, distortion occurs after JPEG2000 compression, especially at low bit-rates. This situation affects the performance of some applications such as change detection, classification, anomaly detection etc. In this paper, in order to overcome compression artifacts, a very deep super-resolution approach is proposed as post-processing after JPEG2000 decompression. In the experiments, scenes of the Onera dataset that is shared in the website of IEEE GRSS DASE are used. Experimental results show that the proposed method not only increases the image quality but also improves the change detection performance.

Automated cutting and splicing of rubberized textile cord ply with a consistent splice overlap width is critical process for tire industry. The splicing process makes the splice overlap by pressing the ply leading end against the ply trailing end at a defined angle. Highly accurate control of the splice overlap width means less material scrap leading to significant cost savings. Accurate measurement and control of the splice overlap width can help tire design engineers to optimize performance properties of a tire, enhance product design, manufacturing, and quality. This paper presents a novel approach to constitute real time measurement and control of the splice overlap width of rubberized textile cord ply based on a discrete scale-space framework with automatic scale selection.

In this paper, an algorithm that improves HO performance in heavily-used routes in order to improve overall cellular network performance and reduce negative impacts on the network is presented. When a user is going to experience HO, the proposed algorithm assesses the signal levels from neighboring cells that are close to the serving cell. After determining the location of base stations (BS) in the test area, the neighboring states of BSs are examined using k-nearest neighbors (KNN) method. The proposed algorithm is set to maximize the quality of service and received power for the current network status. Besides, received signal power data is collected and analyzed systematically for this study. Numerical results show that the proposed algorithm provides 9.63% performance improvement in terms of service quality and power comparison to the current practice.

It is very practical to photograph and store documents using mobile phones. However, it is difficult to obtain a quality document image due to creases on the paper and limitations of the camera pose. These produce geometric distortions and irregular shadows on the document image. The rectification of geometric distortions requires an estimate of the 3D shape of the photographed document. In this study, we introduce a new approach that can estimate the 3D shape of the document using artificial neural networks. Neural network models extract geometric information from the context of the image to create a 3D shape. In addition, an adaptive thresholding algorithm was used to correct lighting-related distortions. Data reflecting actual document conditions were used to train the neural networks. Therefore, in addition to previous studies, the method can be applied to photograph samples which creased in many different ways and photographed from varying perspectives. Comparative experiments show that the method works well.

In this paper, a machine learning based method for antenna scanning type classification in electronic warfare systems is proposed. The proposed method, unlike other methods in the literature, does not require extra resources. In the simulations, it has been observed that radar scanning types can be successfully classified.

In this paper presents a tri-band, monopole, compact microstrip antenna design that can operate in the Xband and Ku-band for fifth generation (5G) communication technology. The proposed antenna operates in the frequency range 10.37 to 10.69 GHz, 11.65 to 11.71 GHz in the X (8-12 GHz) band and in the range 14.14 to 14.20 GHz in the Ku (kurz-under 12-18 GHz) band. The proposed antenna is designed on a RO3210 plate with a dielectric constant (er) of 10.2 and a thickness of 1.27 mm. In this design, circular slots are made on the radiation plane, thereby increasing the gain and bandwidth. The antenna was simulated in the HFSS simulation environment developed by Ansys. Compared with the other microstrip antennas in the literature, it is observed that the bandwidth and gain parameters are superior.

Easily available recent face image/video manipulation techniques and tools are now being utilized to generate highly realistic manipulated videos known as DeepFakes, which can fool face recognition systems and humans. Thus, it is vital to devise precise manipulation detection methods. Despite the progress, existing mechanisms are limited to the datasets or manipulation types. In this paper, to increase the performance under unseen data and manipulations, a DeepFakes detection framework using metric learning and steganalysis rich models is presented. Extensive empirical analysis on three publicly available datasets, namely, FaceForensics++, CelebDF, and DeepFakeTIMIT, were carried out to evaluate the generalization capability of the proposed approach. The framework attained 5% to 15% accuracy gains under unseen manipulations.

Mobile edge computing is a network architecture concept that enables cloud computing capabilities and an information technology service environment at the edge of the cellular network. Basic idea behind is that by running applications and performing related processing tasks closer to the users. Hence, network congestion can be reduced and applications can perform better service to users. In a typical mobile edge computing, virtual machines are allocated to physical machines. Analyzing the number of needed virtual machines and quality of service parameters are important for designing issues as well as cost of the systems. This paper presents an analytical model and analysis of clustered servers in mobile edge computing that can be used to obtain the expected quality of service parameters such as the mean queue length and the throughput. The proposed model considers an open queuing model for analysis. The exact Spectral expansion solution approach is used to obtain performance parameters for the proposed model. The obtained results show that the number of virtual machines is important to achieve the expected quality of service under different circumstances.

In this study, a photometric stereo method, which can model the three dimensional surface topography of documents, is proposed. With the creation of high quality threedimensional model, it is aimed to better analyze the security features, pen tip strokes and other impressions on paper by document examination experts. Compared to 2D document images, modeling the three-dimensional surface topography of documents provides us more sufficient information for solving the order of writing (if two lines partially crossed), detecting the destruction of writings, analyzing the change of pressure, and author determination. Experimental studies show that the proposed method can make the three-dimensional model of the document surface topography precise and successful at micron levels.

28th Signal Processing and Communications Applications Conference, SIU (2020; Gaziantep, Turkey), Constant Envelope (CE) precoding is a transmission technique, which provides efficient power amplification and can be realised with a single radio frequency (RF) chain at the multi-antenna transmitter. In this paper, we study the multi-user massive MIMO broadcast channel with CE precoding and with non-orthogonal multiple access (NOMA). The motivation for this is to provide different multi-user interferance (MUI) energy levels to different users with different MUI requirements. Wepair users according to their channel conditions and examine the impact of the NOMA on MUI. Simulations for the independent and identically distributed Rayleigh fading channel show that different MUI energy levels is possible with NOMA under CE precoding., Sabit genlikli önkodlama, çok antenli ileticide tek bir radyo frekans zinciri kullanarak, etkili ve uygulanabilir güç yükseltimi sağlayan bir iletim tekniğidir. Bu bildiride, yoğun çok girdili çok çıktılı ve çok kullanıcıya iletim yapılan sistemlerde sabit genlikli önkodlama kullanımı incelenmektedir. Kullanıcılara iletim yapılırken dikgen olmayan çoklu erişim kullanılmaktadır. Bu yöntemle, farklı ihtiyaçlara sahip kullanıcılar için farklı girişim enerjisi seviyeleri sağlamak hedeflenmiştir. Kullanıcıların kanal durumları gözetilerek kullanıcı eşleştirme tekniği kullanılmış ve dikgen olmayan çoklu erişimin kullanıcılarda elde edilen girişim enerjisi üzerindeki etkisi incelenmiştir. Bağımsız ve eş dağılıma sahip Rayleigh sönümlü kanallar üzerinde yapılan benzetim sonuçlarına göre dikgen olmayan çoklu erişim kullanıldığında farklı seviyelerde çok kullanıcılı girişim ene

Crop classification is one of the foremost and most challenging applications of remote sensing. Crops exhibit both high intra-class variance across geographical locations, as well as low inter-class variance especially across seasons. As such, they require both spectral and temporal input, both of which are provided by the Sentinel 2 satellites. In this paper, we present the preliminary results of our multispectral and multitemporal crop classification analysis, on a region-wide scale, encompassing multiple climatological conditions and a high number of crop types. We have experimented using the ground-truth provided by the Farmer Registration System, with both well-known spectral and spatial shallow features and classifiers, at both pixel and field level, as well as with state of the art 3D convolutional neural networks. Our results show that Sentinel 2 imagery exhibit a strong potential as input for a systematic crop classification infrastructure.

While airplanes are waiting in the apron, some routines, such as fuel, catering and cargo loading, must be performed before take-off and after landing. During this period images of the airplane in the apron are continuously recorded and evaluated in order to monitor whether these operations are carried out in accordance with the rules without any security problem. Information from different systems and visual evaluations are combined to determine the time, duration and type of operations for each airplane, and then being evaluated for safety and effectiveness of time usage. The aim of this paper is to provide an infrastructure for automatic analysis of aforementioned metric values of an airplane staying in apron. This is done by designing and implementing deep learning instance segmentation algorithm with Mask R-CNN, which is based on detecting the airplane in the video images and by pointing out the components of airplane such as front, back, tail, and gates of the airplane.

In this paper, we demonstrate a study on the use of visible light communications (VLC) for downhole gas pipeline monitoring systems. As per the real field parameters, the model consists of a 7 inch diameter carbon steel coated pipe, a detector, and a light source such that the medium possesses properties of methane gas. Initially, we vary the link distance between the detector and the light source to investigate its effect on detected rays with predefined half viewing angle of the source and active area of the detector. We then investigate the effect of increasing the half viewing angle on the received power for different distance and active detecting area. The channel impulse response also is studied for different link distances (1 m to 10 m). Lastly, we analyze the path loss over multiple distances, and we include it for a benchmark setup along with proposed setup simulations.

In binary classification problems, such as network intrusion detection, class specific observations are unevenly distributed, and the target class (suspicious activity in the network) is rarely observed compared to the non-target (ordinary network operations). Therefore, type I error (prediction of a non-target observation as target) and type II error (failure to detect target) have different consequences. For example, a network security system that generates too many false alarms (type I error), loses its reliability as it will cause rigorous but unnecessary measures to be taken. For this reason, asymmetric class costs should be used in the designed classifiers and the type I error should be controlled directly. To this end, the Neyman-Pearson (NP) classification framework is suitable for binary classification problems under asymmetric error costs, since it minimizes type II error and keeps type I error below a user-specified threshold. In this paper, we propose an NP classification method that solves 1Bu calisma Turkiye Bilimsel ve Teknolojik Arastirma Kurumu (TUBITAK) tarafindan 118E268 sozlesme numarali arastirma projesi kapsaminda desteklenmektedir nonlinear problems via context trees in an online manner. The performance analysis of the proposed method is performed on well-known data sets and the classification behavior is investigated on the basis of ROC (receiver operating characteristics) curves as well as the capability to achieve the desired type I error. According to our analyzes, the proposed algorithm successfully provides an average of 66% increase in the area under the ROC curve along with a precise control over the desired type I error, compared to the algorithms that do not use context trees and can only solve linear problems.

In this paper we inspect two mine detection algorithms [2,3], suggest modifications and present results on detection of anti-personnel (AP) landmines using methods employing Auto Regressive (AR) modeling algortihms. First method is based on the statistical distance between AR models of the reference and simulated data. In literature, while the statistical distance is calculated only for A-Scan data, in this study we suggest statistical distance to be calculated for both A-Scan and rows of the processed data. The second method is relied on AR modeling of the clutter energy in the B-scan. To decide whether a threat signature is present, it is proposed to utilize the difference between the estimated AR model clutter energy and the energy of real data. It is shown that proposed AR model based algorithms can be utilized to detect threat in GPR data and some advices to improve detection performance are given.

In this paper, we propose an automatic text summarization model for Turkish news articles using machine learning models. Our proposed model uses sentence position, speech expression, presence of named entities and statements, term frequency and title similarity as features. We construct and share a new dataset for Turkish text summarization. In our experiments, we show that all our features we use have a positive impact on the performance of the system. In addition, we show that our model outperforms the latent semantic analysis based baseline method.

Irony is defined as the expression of one's meaning by using language that normally signifies the opposite, typically for humorous or emphatic effect. In the textual context, it can be considered as a specific type of opinion mining problem. However, due to the nature of the problem, it is generally more challenging to detect. Irony detection is useful to understand the semantics of a text. It also helps improving the accuracy of many other text mining tasks. In this paper, we study the irony detection problem on Turkish informal texts. In the literature there are several solutions proposed mostly for English texts. However, the number of studies on Turkish texts is very limited. The reason for using informal text is that ironic expressions are seen more frequently in informal communication media such as blogs, social media posts. Previously, performance of conventional supervised learning based solutions, such as Naive Bayes Classifier and SVM, on Turkish texts have been studied. In this work, we investigate the performance of neural network based models in comparison to conventional supervised learning models.

In this paper, we propose a new deep learning based model that uses convolutional neural networks for spatiotemporal crime prediction. To learn the temporal pattern of crime events, we employ a temporally hierarchical structure that branches along the temporal dimension. In addition, channel projection is applied to capture the separate influences of crime events over future crime risk. In the results section, our model is compared with classical methods and the performance is analyzed on publicly available Chicago and Los Angeles crime datasets. The proposed model significantly improves the performance compared to the traditional methods.

Forecasting electricity prices accurately is an essential requirement for the players in the market. The sequential data of electricity poses many challenges such as seasonality and high volatility. Recently, recurrent neural network based methods have showed great success in forecasting time series type problems. In this paper, we propose to use an encoder-decoder recurrent neural network for forecasting hourly electricity prices in the Turkish Day-ahead Market. The approach involves two recurrent neural networks, one to encode the source sequence, called the encoder, and a second to decode the encoded source sequence into the target sequence, called the decoder. We trained and tested our framework on the Turkish electricity price data from 2013 to 2016 and report the accuracy of various recurrent neural network models in terms of mean absolute error. The encoder-decoder recurrent neural networks achieved better accuracy compared to classical recurrent neural networks.

Named Entity Recognition (NER) is a natural language processing task for extracting and classifying important entities such as person, location, and organization. In this paper, we propose a model including methods to lookup word vectors in word embeddings for the Turkish language. We study the proposed methods comprehensively through extensive experiments and out-of-vocabulary analysis. The proposed model only uses word embeddings and character representation via Conditional Random Fields followed by Bidirectional Long Short Memory to tag the sequence of words with the named entity tags. The proposed model achieves 94.05% F1 score on the most commonly used news dataset for the Turkish NER, which is the state of the art in the literature.

Face recognition is a key task of computer vision research that has been employed for various purposes. There are many studies in the literature that attempts for solving this problem. Especially in recent years, deep learning methods have been utilized. With these methods, algorithms can reach higher accuracy levels. In addition to deep learning, the use of representative learning is used to aggregate face images in multidimensional metric spaces. While clustering, the similarities in the faces are generally utilized with the help of Siamese or triplet cost functions. In the studies following this approach, the recognition of the faces is done only by using identity labels. However, in this paper, we propose the method which is utilizing both the identity labels and the attribute labels such as the light intensity of the images and the direction of the face. Therefore, two labels are available for each image. In this study, the images in the Yale-B face dataset are used for training a state-of-the-art architecture, ResNet with the labels which are re-organized by us. In order to benefit from two labels during the training, we employ quadruplet selection methods and a quadratic cost function. The contribution of secondary labels on recognition accuracy was measured in experiments. In addition, the methods for selecting quadruplets are compared.

Earthquakes cause the deadliest and most costly disasters among all natural hazards. Geophysicists and data scientists have spent a lot of effort, trying to predict earthquakes time, location or magnitude to minimize these disastrous effects. However, earthquakes prediction remains a challenging problem. In this paper, the generated aperiodic earthquake failures data by Los Alamos National Laboratory (LANL) were utilized to implement three models: i) random forest (RF), ii) convolutional neural network (CNN) units followed by long short-term memory (LSTM) layers, and iii) hand-crafted features combined with CNN and LSTM layers. Among all, applying a network of CNN and LSTM layers to hand-crafted features is the most accurate and the fastest model to predict the time remaining for the next earthquake. This model achieved the prediction goal with a mean absolute error (MAE) of 1.51 in 44 seconds. These results are promising compared to other models. Thus, we anticipate that our model can be improved and applied to real seismic data that can save thousands of lives and billions of dollars in infrastructure.

Symbolic regression is the process of finding a mathematical formula that fits a specific set of data by searching in different mathematical expressions. This process requires great accuracy in order to reach the correct formula. In this paper, we will present a new method for solving symbolic regression problems based on the firefly algorithm. This method is called Firefly Programming (FP). The results of applying firefly programming algorithm to some symbolic regression benchmark problems will be compared to the results of Genetic Programming (GP) and Artificial Bee Colony Programming (ABCP) methods.

In next generation networks, with the increasing number of diverse mobile network service types, a major challenge lies in how to manage and support all mobile service users who have different Quality of Service (QoS) requirements. Network slicing term can be a solution to satisfy the heterogeneous network requests over a common physical infrastructure. Splitting the network into slices which have different properties (e.g., bandwidth requirements, delay tolerance, user density, etc.) allows to schedule and optimize the requests in constraint of limited resources. The network has to decide to accept or reject the requests, and scale up/down the slices by considering the user density in accepted requests, and then, schedule the accepted requests to serve them in an order. In this paper, effect of scaling up/down the slices by using deep reinforcement learning (DRL) algorithm to the speed of satisfaction of user requests is examined.

The Internet of Things (IoT) is a network of physical devices with embedded technology, such as sensors, RFID and network connectivity, to communicate with other devices or external environments. Such devices have intelligent features for collecting, analyzing and even making decisions without any human interaction. Security is an important requirement in such cases, and authentication is of paramount importance, especially in view of the damage that may result from an unauthenticated malicious device in an IoT system. This paper provides a secure and energy-efficient user authentication protocol proposed to address the above-mentioned issue in the IoT network, which consists of reasonable computing power and a number of resource constrained nodes. A new energy consumption model is presented considering the distance between the devices. With this method, authentication mechanism is combined with the idea of an effective mechanism. The cost of the scheme is analyzed by testing the resistance of the schema to conventional attack types.

The segmentation of cells is necessary for biologists in the morphological statistics for quantitative and qualitative analysis in Phase-contrast Microscopy (PCM) images. In this paper, we address the cell segmentation problem in PCM images. Deep Neural Networks (DNNs) commonly is initialized with weights from a network pre-trained on a large annotated data set like ImageNet have superior performance than those trained from scratch on a small dataset. Here, we demonstrate how encoder-decoder type architectures such as U-Net and Feature Pyramid Network (FPN) can be improved by an alternative encoder which pre-trained on the ImageNet dataset. In particular, our experimental results confirm that the image descriptors from ResNet-18 are highly effective in accurate prediction of the cell boundary and have higher Intersection over Union (IoU) in comparison to the classical U-Net and require fewer training epochs.

In this paper, we propose a novel multi-task variational auto encoder (VAE) based approach for joint source separation and classification. The network uses a probabilistic encoder for each sources to map the input data to latent space. The latent representation is then used by a probabilistic decoder for the two tasks: source separation and source classification. Throughout a variety of experiments performed on various image and audio datasets, source separation performance of our method is as good as the method that performs source separation under source class supervision. In addition, the proposed method does not require the class labels and can predict the labels.

Automated retinal vessel segmentation is useful for diagnosis of many pathological diseases such as diabetic retinopathy, hypertensive retinopathy, and glaucoma. In this paper, an algorithm for semantic segmentation of retinal vessels based on semantic pixel-wise segmentation (SegNet) is presented where the network is initialized using the VGG-16 network. The High-Resolution Fundus (HRF) image database has been used for training and testing the network, where in this dataset 15 images from healthy, 15 images from diabetic, and 15 images from eyes with glaucoma are hand labeled for retinal vessels. Because of the limited number of hand-labeled images, first, a data augmentation has been applied where reflection, scaling, translation, and rotation are used for this purpose. Then class weighting is applied to minimize the issues due to the imbalanced dataset. When tested with HRF dataset, sensitivity has been estimated as 73.72%, specificity has been estimated as 94.85%, and accuracy has been estimated as 93.23%.

Designing an hardware architectiire using a lowlevel hardware description language (Verilog, VHDL) is a difficult and time-consuming task especially when the application is a complex and memory intensive one. A high-level synthesis (HLS) tool, most recently and actively being researched in several research groups, automatically generates an RTL description of the hardware architecture from a high-level (C/C++) program. However, application to be accelerated on the hardware via an HLS tool needs to be written in order to decrease the overall memory access latency by simply rewriting the code so that the reformatted loop structure will have more locality. In this paper, we present two case studies to decrease the memory access latency by improving the locality of a given application by reorganizing the memory access pattern of the application being accelerated via HLS on hardware that has a cache. We also emphasize the importance of locality in performance of hardware accelerators with our empirical results on a Zynq-based SoC platform.

Atmospheric analyses and simulations requires high resolution physical models on super computers that consume many hours of computations. Deep learning and machine learning methods used in forecasting revealed new solutions in this area. Main goal of this paper to solve high-resolution numeric forecasting problem. We present a model architecture for spatiotemporal prediction. Model is composed of Convolutional Longshort Term Memory (Conv-LSTM) units with encoder-decoder structure. Model takes sequence of inputs and outputs the next time sequence. Model is trained in a supervised manner. Experiments are made on high-scale benchmark numerical weather dataset. All selected baseline models are surpassed by 3°C mean square error (MSE) with statistically significant results. Both spatial and temporal changes in the temperature is captured. Model forecasted 5 time steps with 1, 3 and 24 time difference successfully.

The security of lattice-based cryptosystems in postquantum cryptography is mainly based on the difficulty of solving the shortest vector problem (SVP) or the closest vector problem (CVP). In this paper, efficient implementations of GaussSieve and ProGaussSieve algorithms alternatives to solve the shortest vector problem are studied. A cryptographic software library is developed which is an infrastructure for the implementation of sieving algorithms in order to have faster running times than the ones in the literature. In addition, a comparison table including the running times of the implementations of GaussSieve and ProGaussSieve algorithms is presented. According to the experimental results, the proposed GaussSieve and ProGaussSieve implementations are much more faster and efficient than the others.

This paper proposes a new hybrid approach for solving gene selection problems in cancer microarray data, which is one of the most challenging tasks in bioinformatics. Minimum-redundancy-maximum-relevance (mRMR) filter approach is combined with the binary black hole optimization algorithm (BBHA) to pick out extremely discriminative genes from cancer datasets. The support vector machine (SVM) is employed as a fitness function to accurately diagnose cancer. The experimental results prove that the suggested method exhibits better classification accuracy with the smallest gene subset compared to existing state-of-art methods.

A wireless multi-hop network is a collection of nodes that communicate by message passing over multiple links. Sending a message to a remote node can consume some energy from all intermediary nodes. In a network, the nodes with minimum distance to all other nodes are called Jordan central nodes. Selecting the central nodes as sink or base station can considerably reduce the overall energy consumption and increase the network life time. This paper proposes a new parallel algorithm to find all central nodes of a network by finding BFS trees of a subset of nodes. The roots of the trees with smallest height are selected as Jordan central nodes. After finding each tree the algorithm eliminates some nodes from the search space. Available processors construct the BFS tree for different nodes in parallel and eliminate a group of unvisited nodes after creating each tree. The implementation results of the algorithm using different number of processors on topologies with up to 250 nodes showed that the proposed algorithm can find all central nodes by examining less than 20% of nodes in less than 0.034 seconds.

Arrhythmia is any irregularity of heart rate that cause an abnormality in your heart rhythm. Manual analysis of Electrocardiogram (ECG) signal is not enough for quickly identify abnormalities in the heart rhythm. This paper proposes a deep learning approach for detection of five different arrhythmia types based on 2D convolutional neural networks (CNN) architecture. ECG signals were obtained from MIT-BIH arrhythmia database. For CNN architecture, each ECG signal was segmented into heartbeats, then each heartbeat was transformed into 2D grayscale heartbeat image. 2D CNN model was used due to success of image recognition. The proposed model result demonstrate that CNN and ECG image formation give highest result when classified different types of ECG arrhythmic signals.

This paper is intended for a Defected Ground Structure (DGS) based miniaturization of a circular patch antenna for 5G mobile applications. From a standard circular patch antenna radiating at 3.6GHz, it was possible to obtain 50 and 70% size reduction by the help of DGS. As a result, 75 and 91% area shrinkage were achieved respectively. The antenna is made of 0.762 mm thick Arlon DiClad 880 substrate with 2.2 relative permittivity and a 60×60×0.035 mm3 metallic ground plane. All simulations were performed using CST Studio Suite. The achieved gains are as follows: 8.69dB for the standard antenna, 5.17 and 2.41dB for 75 and 91% reduced area antennas respectively. The reflection coefficients are −40.03dB and −32.85dB for 75 and 91 % reduced area antenna respectively, while the standard antenna's reflection coefficient is −46.59dB. The obtained efficiencies are 91% for reference antenna, 99% and 95% for 75 and 91% reduced area antennas respectively. The miniaturized antennas provide a desirable bandwidth, covering 3.4–3.8 GHz band, suitable for 5G mobile communication.

Convolutional Neural Networks can solve the target detection problem satisfactorily. However, the proposed solutions generally require deep networks and hence, are inefficient when it comes to utilising them on performance-limited systems. In this paper, we study the infrared target detection problem using a shallow network solution, accordingly its implementation on a performance limited system. Using a dataset comprising real and simulated infrared scenes; it is observed that, when trained with the correct training strategy, shallow networks can provide satisfactory performance, even with scale-invariance capability.

Segmentation networks have proven to be popular tools for large scale pixel-wise remote sensing image classification as they can deal with wide spatial areas efficiently, as opposed to convolutional neural networks trained with pixel centered patches. However, they are often criticized in terms of spatial consistency. As such, they have received various extensions through the last few years, in the form of dilated convolutions and skip connections and more. In this paper, we address the same issue by feeding attribute filtered images, that contain inherently a multiscale hierarchical representation of the underlying image, as input to a segmentation network, in an effort to both accelerate convergence and render easier the feature learning task of the bottom layers. We validate our approach through the production of land-use and land-cover maps for a large area of Turkey using Sentinel 2 multispectral images and ground truth from the Copernicus Land Monitoring Service.

In recent years, with the rapid increase of voice data, solutions are being sought extensively for examination and indexing in the field of speech processing. One of these solutions is the speaker diarization, which is used to examine speech records that include multi-speaker. The speaker diarization system basically splits the speech file into segments using the speech file's silence fields and examines the similarity between the segments. In this study, deep learning based embedding vectors are used for speaker representation. The vector embedding proposed for speaker representation is performed with x-vectors extracted using time-delayed deep neural network and d-vectors extracted using LSTM. Then the system is tested with xd-vectors consisting of the combination of these two vectors. As a result, the effect of representative embedding vectors on the performance of the diarization system is examined in the scope of this paper.

The assessment of uterine contraction provides valuable information about the progress of labour. Deliveries occurring before the estimated date introduce undesirable consequences for the fetus and mother. Electrohysterography (EHG), a noninvasive monitoring method, is the most preferred and studied method to diagnose preterm delivery and minimize its unpleasant consequences. This paper proposes a novel feature extraction method for non-invasive diagnosis of preterm delivery using EHG signals. The acquired EHG signals are sliced into frames for spectral analysis. Next, the centroid frequencies of the frames are computed to establish a feature vector. Using a 1D Convolutional Neural Network (CNN), the effectiveness of the feature vector is compared with existing methods in the literature. The experimental results confirm that the proposed method shows superior performance and can be used to predict term-preterm birth more accurately.

With the rapid development of Smart Grids over wireless sensor networks, the reliable image and video transmission has become more imperative and faces many challenges in wireless harsh environment. Transmission of the image data may experience distortions that decrease the quality of the image due to noise, interference and path loss. In this paper, we show the effect of two error correction algorithms (Static and Adaptive RS) in order to increase the perceptual quality of the transmitted images. Here, as Static Reed-Solomon (RS) coding utilizes additional bits to the information transmitted to correct errors at the receiver side. Unlike Static RS, Adaptive RS is a derived version of the Reed-Solomon that sets the error correction codes of the algorithm based on the distance between the nodes adaptively. Performance results obtained from numerous simulations indicate that the Adaptive RS exhibits much better performance results in terms of PSNR, BER, and Throughput than Static RS. It is good candidate for providing application layer perceptual quality requirements for Smart Grid applications.

In this paper, we present a Unet architecture made of octave convolution for dental image segmentation problem. In this architecture, the requirements for memory and accuracy are significantly improved compared to previous works in the literature. Compare to state-of-art models on this topic the classification accuracy in dental image segmentation is increased by %2, and the memory usage is decreased by %70. Suggested architecture showed a performance of success on ISBI2015 dataset.

This paper provides new results over our previous work presented in ICIP 2019 on the performance of learned frame prediction architectures and associated training methods. More specifically, we show that using an end-to-end residual connection in the fully convolutional neural network (FCNN) provides improved performance. In order to provide comparative results, we trained a residual FCNN, a convolutional RNN (CRNN), and a convolutional long-short term memory (CLSTM) network for next frame prediction using the mean square loss. We performed both stateless and stateful training for recurrent networks. Experimental results show that the residual FCNN architecture performs the best in terms of peak signal to noise ratio (PSNR) at the expense of higher training and test (inference) computational complexity. The CRNN can be stably and efficiently trained using the stateful truncated backpropagation through time procedure, and requires an order of magnitude less inference runtime to achieve an acceptable performance in near real-time. / Bu bildiri, ICIP 2019’da sundugumuz öğrenilmiş video çerçeve öngörü mimarileri ve egitim yöntemleri karşılaştırılması çalışmamız üzerine yeni sonuçlar vermektedir. Daha spesifik olarak, tamamen evrisimsel sinir agında (FCNN) uçtan uca kalıntı bağlantısı kullanarak öngörü performansında gelişim sağlandığını gösteriyoruz. Sonuçları karşılaştırmak için, ortalama kare kaybını kullanarak sonraki kareyi öngörmek için kalıntı baglantılı FCNN, evrisimsel bir RNN (CRNN) ve evrisimsel uzun-kısa süreli bellek (CLSTM) agı eğittik. Yinelemeli ağlar için durum bilgisi taşımayan (stateless) ve taşıyan (stateful) egitim gerçekleştirdik. Deneysel sonuçlar, kalıntı baglantılı FCNN mimarisinin yüksek egitim ve test (çıkarım) hesaplama karmaşıklığı pahasına PSNR açısından en iyi performansı verdigini göstermektedir. CRNN, durum bilgisi taşıyan zamanda geri yayılım yoluyla kararlı olarak egitilebilmekte ve kabul edilebilir bir performansla gerçek-zamanlıya yakın çerçeve öngörüsü yapabilmektedir., Scientific and Technological Research Council of Turkey (TÜBİTAK); Turkish Academy of Sciences (TUBA)

The detection of an emitting radar and accurate tracking of the associated platform direction are the most important goals of Electronic warfare (EW) systems. EW systems use Kalman filter in order to compansate the measurement errors that occur during instantaneous direction measurements of detected radars. The compatability of the Kalman filter prediction model and parameters with the platform motion type improves the quality of the values shown on the screen. In this paper, it is aimed to determine the platform types associated with the radars using radar parameters that are measured by EW systems. First, an efficient pre-processing method including quantization for interval values and grouping of class values is applied to improve classification performance. Then, Multi-task Learning (MTL) neural network and Support Vector Machine (SVM) techniques are applied to the problem. It is observed that SVM technique outperforms MTL technique. It is evaluated that the result of the classification process can be utilized for selecting the type and the parameters of the filter used for tracking.

Innocent social media users getting mobbed by cyberbullies is a common situation today. As the number of internet users increases, number of mobbing cases increase rapidly as well. That is why the detection of messages that contain cyberbullying becomes an important problem to be solved and the current solutions like reporting mechanism which is used by most of the social media platforms are not sufficient enough. While there are many studies on the detection of cyberbullying in the English language, there are only a few studies in Turkish. In this study, convolutional neural network (CNN) models were used for the first time for cyberbullying detection problem in Turkish language. Different pretrained word embeddings have been used as input to the CNN model and their effects have been analysed. Finally, this paper has achieved the highest performance so far on Turkish cyberbully detection task with 0.937 F1 score.

Deep Reinforcement Learning (DRL) algorithms are used in many challenging tasks and their usage areas are rapidly increasing. One of these areas is the formation flights of Unmanned Aerial Vehicles (UAVs). The rising of Reinforcement Learning (RL) algorithms performances is directly proportional to the development of environments. This paper presents a new environment developed through software (Ardupilot, Mavlink, drone-kit) that is frequently used in open source UAV simulation and programming, and the performance of the Evolutionary Reinforcement Learning (ERL) agent in this environment. The difference of this environment is that, unlike other environments, the model can be operated directly on a drone-kit supported vehicle and is specifically defined on the centralised formation task. The aim of this study is; in order to question the performance of the Evolutionary Reinforcement Learning (ERL) algorithm which has better results than other algorithms in DRL training environments,in this environment, and increasing the usage of the algorithm in this direction.

In this paper, a study on semantic segmentation of the spheric (healthy) and aspheric (pathological) femoral heads in magnetic resonance (MR) slices with the deep Convolutional Neural Networks (CNNs) is presented. Femoral heads in bilateral hip MR slices were successfully segmented with U-NET is an encoder-decoder network based deep learning architecture. In the proposed study, bilateral hip MR slices which were acquired with different imaging protocols in coronal imaging plane of the patients diagnosed with Legg-Calve-Perthes (LCP) disease were used. In experimental studies, quite successful results have been achieved on a small amount of MR image data in segmentation of the healthy and pathological femoral heads. Performance tests evaluated on a total of 66 femoral head images in 33 MR slices of 13 LCP patients show that proposed study has a segmentation accuracy of approximately 89%.