Search

In this study, it is aimed to produce meaningful words in accordance with character-based Turkish grammar rules by using encoder-decoder and attention architecture, which are deep learning algorithms. The results of the developed model are compared with the results of LSTM and GRU models, which are other deep learning algorithms. It is seen that the language models created with LSTM and GRU models give similar results at 100 and 200 epoch values and at different threshold values of the temperature sampling method. Among these models, the GRU model gives the highest success value with 88.40% at 200 epochs and 0.5 temperature threshold value. The encoder-decoder and attention language model developed for this study gives the highest success value of 91.90% at 100 and 200 epoch values and at different threshold values of the temperature sampling method at 200 epoch and 0.5 temperature threshold value. At the end of the experiments, the encoder-decoder and attention architecture model showed an average of 2.83% more success than the LSTM model and an average of 0.19% more success than the GRU model. [ABSTRACT FROM AUTHOR]

Firms in the manufacturing sector need to constantly monitor their performance in order to maintain their development under competitive conditions in the market. In this study, eleven KPIs are determined to measure the production performance by taking into account the factory assets. The proposed system is designed in which the relevant KPIs are obtained via the instantaneous data received from the CNC machines in a production system. The main objective of this study is to measure and estimate production performance. In this way, it is aimed to provide a proactive approach for the assets whose performance is monitored by the decision-makers. LSTM and LightGBM models which are deep learning techniques are proposed for the estimation of performance indicators. The approximately three-month time series OEE (Overall Equipment Effectiveness) values of the sample CNC machine are used for estimation. The estimation performance of methods is compared over performance metrics (MSE, MAE, etc.). The results indicated that LightGBM outperforms LSTM for all performance metrics. [ABSTRACT FROM AUTHOR]

Thanks to the precision agriculture technologies developed in recent years, many processes such as irrigation, fertilization, spraying, weeding and harvesting of agricultural products can be done by autonomous systems. Especially in some plant species such as sunflower, when to apply these processes is largely decided according to the developmental stage of the plant. In this study, deep learning methods were used to classify the developmental stages of sunflower plants. Since the images taken with the drone are high resolution, each of them is divided into 6 equal parts, and then 8 classes are determined and the images belonging to each class are extracted. A data set consisting of 12800 images in total, 1600 in each class, was created. Six different deep learning models, namely AlexNet, InceptionV3, ResNet101, DenseNet121, MobileNet and Xception, were tested with Sgd, Adam and Rmsprop optimization methods and their performances were compared. In order to evaluate the success of the models correctly, the trained models were also tested on a second data set created with images taken from a different terrain and high success rates were obtained. In addition, a 7-class test set was created for images that could not be clearly determined at which stage the plant was in, and the success rates of the models were tested. It was observed that the success rate was very low for the images in the 7-8 intermediate class, and the filters used in the image processing techniques that would increase the success rate for this class were applied to the images, and the models were retrained and the results were evaluated. [ABSTRACT FROM AUTHOR]

After lung cancer, prostate cancer is the most common type of cancer in men. Nowadays, advanced prostate imaging is conducted by multiparametric prostate magnetic resonance imaging (MRI) performed by radiologists. Although many pathologies of the prostate can be visualized, the main purpose is to determine the probability of prostate cancer and to decide whether a biopsy is needed. This process includes the analysis of different series of Magnetic Resonance (MR) images, including T2-weighted images (T2W), diffuse-weighted images (DWI), and dynamic contrast enhanced images (DCE). However, firstly it is necessary to differentiate the prostate zones. Then, lesion scanning is performed in the relevant prostate zones. Finally, there is a need to perform prostate lesion scoring according to PI-RADS v2. Therefore, diagnosis of prostate cancer is a complex and long process. For this reason, decision support systems are needed for the diagnosis of prostate cancer. In this context, the main purpose of the study is to provide automatic segmentation of prostate regions. In the segmentation task, Mask R-CNN algorithm was used with T2W MR images of 15 patients. With the use of the ResNet-50 backbone model, the Mask R-CNN algorithm achieved the segmentation performance with a mAP50 value of 96.040. Finally, the trained model was integrated into the PACS system. Thanks to the integration, a ready-to-use in hospitals artificial intelligence-supported decision support system was developed. Thus, time was saved while reducing the workload on health employees. [ABSTRACT FROM AUTHOR]

It has become imperative to monitor the data-driven industrial systems of today's technology before potential failures occur. Predictive maintenance predicts these failures before they occur and takes the necessary action to prevent malfunctions from occurring. In this study a comparative predictive maintenance application which is based on machine and deep learning is realized. Logistic Regression, Naive Bayes Classifier, Decision Tree, Support Vector Machine, Random Forest, and K-Nearest Neighborhood are used as the classical machine learning methods while Long Short-Term Memory and Gated Recurrent Unit are used as the deep learning architectures. The performances of the methods are examined on the Predictive Maintenance dataset from UCI Machine Learning Repository for fault type detection and the results are presented comparatively in terms of metrics in detail. In the experimental studies, fault type detection is handled separately in the form of multiple and binary classification problems. In the solution of the multiclassification problem, the highest accuracy among the machine learning methods is obtained by Random Forest method with 98.26%, while the accuracy value obtained with both deep learning architectures is 97.51%. In the solution of the binary classification problem, after the data balancing, the highest accuracy among the machine learning methods is obtained by Random Forest method with 95.03%, while the highest accuracy among the deep learning architectures is obtained by Gated Recurrent Unit architecture as 93.03%. [ABSTRACT FROM AUTHOR]

Nowadays, the Internet is a structure that people can access easily and at the same time produce content easily and without control. In parallel with this situation, the ability of extract information from the raw data that makes up big data has become more complex. The fact that the headlines of the contents contain uncontrolled and misleading elements makes it difficult to reach the right information. The headlines of the contents are important for people to reach the information they want in their limited time. In this study, it is aimed to produce headlines suitable for the content instead of headlines that may be misleading for news. For this purpose, an application that produces headlines for Turkish news with deep learning method has been developed. SuDer news corpus is used as dataset. For the training of the model, it is aimed to obtain more humanoid results in the production of news headlines by using the Transformer architecture, which is frequently preferred in natural language studies today and the abstract summarization method. In this study, in order to compare the performance of the Transformer model, models are prepared and trained with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures. At the end of 25 epochs of training with LSTM, GRU and Transformer architectures on the corpus, the values of loss are 1.03, 0.55 and 2.49 respectively. In the experiments performed on the validation data, measurements are made with ROUGE-1, ROUGE-2 and ROUGE-L metrics. As a result of the measurements, it is observed that the Transformer architecture is partially good, based on the metric values produced. In addition, when the headlines produced with these architectures are examined, it is observed that the headline obtained with the Transformer architecture produce headlines that are partially more suitable for the news content compared to other architectures. [ABSTRACT FROM AUTHOR]

In this study, the Deep Learning technique was combined with the Uncertainty Quantification scale to determine whether patients have cancer or not, and if cancer is present, to classify it as benign or malignant using a dataset of 663,549 lung tomography data. The reliability of this model was also investigated. 80% of the mentioned images were used for training, 10% for testing, and 10% for validation. Two methods were proposed for this purpose. In the first method, measurements were made without using Uncertainty Quantification, and the accuracy rate for the ResNet50 model was found to be 0.965, while the accuracy rate for the AlexNet model was calculated as 0.837. Additionally, the accuracy rate for the Inception v3 model was found to be 0.849, and for the VGG16 model, it was 0.826. Firstly, the accuracy rate of the 4-Layer Convolutional Neural Network model was calculated as 0.971. In the second method, measurements obtained using Uncertainty Quantification (UQ) resulted in an accuracy rate of 0.89 for the Bayesian Neural Networks ResNet50 model. Furthermore, the accuracy rate for the Bayesian Neural Networks AlexNet model was 0.86, for the Inception v3 model it was 0.91, and for the VGG16 model, it was 0.77. The study has innovative features such as the development of a new model called the 4-Layer Convolutional Neural Network and the use of uncertainty quantification methods. These results demonstrate that the performance of specific neural network models can be enhanced by considering uncertainty. [ABSTRACT FROM AUTHOR]

In this paper, we present a web-based optical character recognition (OCR) system that converts images of Ottoman documents printed with naskh font into text using CNN+RNN-based deep neural network models. For training, three datasets - original, synthetic, and hybrid - were prepared and three different OCR models were created. The original data set consists of 1,000 pages and the synthetic data set consists of 23,000 pages. Hybrid data set contains both. The trained models were compared with Tesseract's Arabic and Persian, Google Docs' Arabic, Abby FineReader's Arabic, and Miletos OCR model/tools with a 21-page test set. The comparison was made with 3 different texts (raw, normalized, and joined) and using 3 different criteria (character, ligature, and word recognition). The Osmanlica.com Hybrid model produced significantly better results than the others with 88.86% raw, 96.12% normalized, and 97.37% joined accuracy in character recognition; 80.48% raw, 91.60% normalized, and 97.37% joined accuracy in ligature recognition; and 44.08% raw and 66.45% normalized accuracy in word recognition. To investigate the effects of the characteristics of the alphabet on OCR, character, ligature, and word frequency analyses of Ottoman was performed. In this analysis, the characters in the alphabet were grouped according to distinctive features such as connectedness, letter body, position and number of dots, type of character, and source language; and frequencies and recognition accuracies were examined for each group. OCR results are also reported for each character. [ABSTRACT FROM AUTHOR]

There are about 466 million hearing and speech disabilities in the world, and there are about 1.5 million hearing and speech disabilities in Turkey. Sign language (SL) is the language that is used by hearing and speech-impaired individuals to communicate among themselves and with other individuals. This language realized with hand, face, and body gestures is a dynamic language that can vary from country to country and from region to region in the countries where it is spoken. Learning sign language at a younger age is important, especially in terms of both cognitive and intellectual development and the academic development of the individual. For this reason, there are a series of studies developed to use technology in sign language learning. The conducted studies reported the successful results of humanoid robots developed in sign language learning. It has been proven that the developed humanoid robots are successfully performed in sign language learning. There is no developed humanoid robot for sign language learning in Türkiye. On the other hand, the current studies in the literature have been carried out with a limited number of ready-made humanoid robots including Turkish sign language expressions. With the humanoid robot we developed within the scope of the study, some Turkish sign language statements were successfully realized. [ABSTRACT FROM AUTHOR]

Fire is a natural disaster that causes ecological, social and economic damage. With global warming and the widespread use of explosive/flammable chemicals, fires have become one of the most important problems for humanity. Early fire detection is critical to minimize destruction. For this reason, studies on fire detection through images have begun to be carried out in order to detect fires early. In recent studies for fire detection from images deep learning algorithms are generally used. These studies focus on the analysis of images taken from cameras with models running on powerful servers. With the developments in mobile devices and internet of things, images can analyzed on edge devices. In the study fire and smoke detection model which requires low processing power has been developed to enable the images to be analyzed on mobile device without transferring them to server. The MobileNet convolutional neural network was revised, the last 3 layers were removed and replaced with a flattening layer and a dense layer consisting of two nodes. Fire and smoke detection model, the method with the highest accuracy was determined among the models developed using revised MobileNet, original MobileNet, MobileNetV2, EfficientNetB0, ShuffleNet, NASNetMobile and PeleeNet convolutional neural networks. To train and test the models, 80% of the dataset consisting of 43,355 images was used for training and 20% for testing. According to test results, highest accuracy rate of 98.37% was achieved in the revised MobileNet. [ABSTRACT FROM AUTHOR]

Patents are documents that help protect innovations in information technologies and grant special rights to the creator of these innovations for a certain period of time. While these rights give the patent owner the right to use the innovation commercially, they prevent others from using the innovation without permission. Radical innovations and ground-breaking technological advances are derived from technical information contained in existing patents. Using an automatic classification system, patents assigned to the technical class to which they belong can pave the way for researchers and provide an environment in which they can create new inventions. This study presents an automatic patent classification analysis using the BERT algorithm. Hyperparameter analyses are also preferred in this study in order to achieve more successful prediction accuracy in automatic patent classification problems. The obtained results were at a level that competed with those in the literature. An accuracy of 58% was achieved at the subclass level. [ABSTRACT FROM AUTHOR]

To fully utilize the potential of solar energy and effectively operate solar energy systems, it is vital to know solar radiation values. Modeling solar radiation data with high variability is a complex problem, and nonlinear methods are needed. This study proposes a hybrid model using CNN and LSTM architectures for solar radiation prediction. The performance and applicability of the proposed model are examined by comparing it with ARIMA statistical method and different machine learning methods such as Random Forest, Decision Tree and K-Nearest Neighbor. The study used hourly solar radiation values measured with a pyranometer on the Afyon Kocatepe University campus. As a result of the experimental studies, the CNNLSTM model is the most successful model in terms of RMSE, MAE, MAPE and r² evaluation criteria, achieving 16.92%, 17.25%, 17.24%, 22.89% more successful results than the Decision Trees model, which produced the most unsuccessful predictions, respectively. [ABSTRACT FROM AUTHOR]

MR spectroscopy (MRS) is a critical imaging modality in brain tumor detection, in which brain metabolites other than water and fat can be measured. Along with the advances in artificial intelligence and data mining, it is seen that computer aided diagnosis (CAD) systems are widely used in medicine. The high achievements achieved have a positive effect on medical diagnosis and treatment processes by providing time saving for CAD applications. Deep learning, based on classical classification methods, the results are more successful on the data, and with high classification ability is the artificial intelligence algorithm. In this study, it was aimed to develop a CAD system to assist physicians in tumor detection by using the ratios of Choline (Cho), creatine (Cr) and N-Acetyl Aspartate (Naa) metabolites (Cho/Cr, Cho/Naa and Naa/Cr) obtained by MRS. For classification data, four classical data classification methods were used, which are deep learning application, Long Short-Term Memory (LSTM) and comparison. While the class performance results obtained with K-Nearest Neighborhood (KEYK), Support Vector Machine (SVM) and Decision Trees (DT) were found to be around 83±6,36 % from classical methods, this value was obtained as 93% with the LSTM-based deep learning method proposed in the study. This shows that deep learning can be used in a CAD system that can assist the experts in the relevant field in the decision-making process in classifying MRS data by demonstrating high classification performance. [ABSTRACT FROM AUTHOR]

Hyperspectral images (HSI) are 3-dimensional (3D) image cubes with two spatial and one spectral dimensions. The development of deep learning methods has had a significant impact on HSI classification. Especially convolutional neural network (CNN) based methods are getting more attention in this field. In this study, we make use of the deep learning architectures LeNet5, AlexNet, VGG16, GoogleNet and ResNet50, which are among the successful examples of CNN for the HSI classification problem. We use a 3D CNN-based hybrid approach when using these architectures. Because, using 3D CNN, spectral-spatial features are extracted simultaneously. In this case, the classification accuracy of HSIs is increased with the spectral-spatial-based deep learning architecture. However, in the proposed model, principal component analysis (PCA) is used as a preprocessing technique for optimal band extraction from HSIs. After applying PCA, 3D cubes are obtained by neighborhood extraction and given to the input of deep learning architectures. Indian pines, Salinas, Botswana and HyRANK-Loukia datasets were used to compare the classification performances of 3D CNN-based deep learning architectures. As a result of the applications, the best classification accuracy was obtained with VGG16 architectures in Indian pines dataset, ResNet50 in Botswana dataset, VGG16 in HyRANK-Loukia dataset, LeNet5 and VGG16 architectures in Salinas dataset. [ABSTRACT FROM AUTHOR]

In parallel with the development of Internet and wireless communication technologies, studies in the field of IoT are also progressing. With the IoT sensors used in the healthcare field, it becomes easier to follow the patients closely. In addition, it is possible to create statistics that will support the treatment process with the data collected from the patients. However, besides the provided facilities, IoT devices that communicate wirelessly and are connected to the internet also have some problems such as security. Serious problems can occur as a result of attacks on different layers of IoT, which is used in healthcare environments. Exposure of sensitive data in the field of Health to any of these attacks can have negative consequences, such as changing the data out of reach of authorized users or capturing it by an attacker. This article attempts to predict duplication, interception, and modification attacks in Message Queuing Telemetry Transport (MQTT) message using artificial intelligence techniques using a complete and labeled IoT dataset containing real world behaviors in IoT networks. On the dataset used, SVM algorithm has Accuracy 85%, f1 98%, Recall 100% values; Naive Bayes (NB) algorithm has Accuracy 89%, f1 86%, Recall 100% values; LSTM has Loss 6.7%, Accuracy 98%, f1 98%, Recall 98%. The LSTM algorithm, which is a deep learning algorithm in detecting abnormal behaviors, has performed better than existing machine learning approaches with low loss and high accuracy data. [ABSTRACT FROM AUTHOR]

Mobile devices, which are becoming more and more widespread today, have turned into hand-held computers thanks to the multimedia communication and applications. Today, the multimedia applications have been supported by traditional mobile phones. With their increasing functionality, mobile devices are used for many purposes such as enriched internet experience, financial transactions, access to social media platforms, sharing, music and video. Performing transactions where sensitive personal data transfers such as banking and shopping are carried out on mobile devices make mobile devices the target of attackers. In this study, a deep learning based malware detection system has been developed based on the interactions of mobile applications on the network. The developed LSTM-based deep learning model has been analysed comparatively with NB, RF, SVM, MLP, CNN, RNN and GRU using accuracy, precision, recall and F-1 metrics. The experimental results showed that the developed LSTM-based deep learning model is more successful in malware detection than others with 95% accuracy. [ABSTRACT FROM AUTHOR]

Driver drowsiness detection is an important issue to prevent traffic accidents. 40% of severe traffic accidents are due to drowsiness. Various methods are used for driver drowsiness detection. One of the driver drowsiness detection method is driver drowsiness detection based on the analysis of signals such as EEG and ECG. Another driver drowsiness detection method is driver drowsiness detection based on vehicle-driver interaction. The last driver drowsiness detection method used in the study is driver drowsiness detection from images. This method is more advantageous than the other two methods in terms of cost and usability because no driver intervention required. Classical image processing techniques and deep learning algorithms are used for driver drowsiness detection from images. Recent driver drowsiness detection studies are based on deep learning models. In addition, the model to be developed will need to be able to work on mobile devices in order to ensure widespread use. In the study, Convolutional Neural Networks were used for driver drowsiness detection on mobile devices. In order to increase the success rate of the model, the pre-trained model was reused with the transfer learning technique. The model developed for the training consists of 14 layers and 1,236,217 parameters. The dataset consists of two categories, open-eye and closed-eye images. The developed model achieved 95.65% accuracy, 95.86% precision, 94.32% recall and 95.17% f1 score which achieved better results than previous studies. A dataset of 2425 images was used to train the model. [ABSTRACT FROM AUTHOR]

Nowadays, the use of the human ear images gains importance for the sustainability of biometric authorization and surveillance systems. Contemporary studies show that such processes can be done semi-automatically or fully automatically, instead of being done manually. Due to the fact that deep learning uses abstract features (i.e., representation learning), it reaches quite high performance values compared to classical methods. In our study, a synergistic gender recognition approach based on hybrid deep learning was created based on the use of human ear images in classifying people fully automatically according to their gender. By means of hybridization, hybrid deep neural network architectural models are used, which include both convolutional neural network component and recurrent neural network type components together. In these models, long-short term memory and gated recurrent unit are taken as recurrent neural network type components. Thanks to these components, the hybrid model extracts the relational dependencies between the pixel regions in the image very well. On account of this synergistic approach, the gender classification accuracy of hybrid models is higher than the standalone convolutional neural network model in our study. Two different image datasets with gender marking were used in our experiments. The reliability of the experimental results has been proven by objective metrics. In the conducted experiments, the highest values in gender recognition with hybrid models were obtained with the test accuracy of 85.16% for the EarVN dataset and 87.61% for the WPUT dataset, respectively. Discussion and conclusions are included in the last section of our study. [ABSTRACT FROM AUTHOR]

The subject of automatic detection of the presence of cervical cancer by evaluating histopathological Pap-Smear images with computerized diagnosis / detection software is an active field of study. The reason for this is that the objects (cell nucleus, cytoplasm, white blood cell, bacilli and speckle) in the visuals overlap and change the geometric structure and pattern of each other, they are dispersed in the image with different density, and the noise patterns are different. In addition, the difficulty and costs of creating a tagged large dataset prevented the emergence of a common dataset in this area. The mentioned difficulties negatively affect the achievements in current classification studies and trigger the need for new approaches. In this paper, a three-step approach based on building large Pap-Smear datasets using Generative Adversarial Networks (GANs) is proposed. Accordingly, in the first step, geometric shape and pattern models of each object structure in Pap-Smear images are created. In the second stage, synthetic Pap-Smear images (Ground True) with the desired number and distribution of objects are produced using the produced parametric models. In the third stage, the performances of existing GANs (Pix2Pix, CycleGAN, DiscoGAN and AttentionGAN) to produce GT are evaluated and a solution-oriented new current GAN architecture (Pix2PixSSIM) is proposed. Experimental studies show that a large Pap-Smear data set can be produced in a very short time with the proposed GAN architecture. In this way, it is seen that deep networks with high classification success can be trained. [ABSTRACT FROM AUTHOR]

Electronic retailing (e-retail) is a type of retailing where the retailer and customers communicate via web and mobile applications. In order for e-retail businesses to continue to grow in an increasingly competitive environment and to stand out from competing businesses, they need to be able to respond to changing consumer demands in a timely and accurate manner. A demand forecasting study using Machine Learning and Deep Learning algorithms for the e-retail sector with local supermarkets' data in Turkey (as far as we know) has not been carried out yet. In this study, the demand for a certain category of products was estimated by considering the two-year e-commerce data (website and mobile application) of a local supermarket for the years 2019-2020 and the factors affecting product sales (holidays, CPI value and unemployment rate). Twenty-four different methods of six different artificial intelligence algorithms (Deep Learning, Artificial Neural Networks, Gaussian Process Regression, Regression Tree, Support Vector Regression and Ensemble Learning) were used to obtain the best demand forecasting model. The obtained results were evaluated using correlation coefficient (R), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) criteria. The best result was obtained using Long Short-Term Memory Networks (RMSE= 0.0353; MAE= 0.0164; R = 0.9742). The results obtained will be able to increase the success of e-retail sales by using the product supply in the right quantities, in sales campaigns, and in marketing strategies. [ABSTRACT FROM AUTHOR]

Character recognition in natural images is a very difficult problem due to many factors such as variability of light, background clutter, severe blur, inconsistent resolution and different scale depth. In addition to these features, distortions in characters and numbers are encountered in street view photographs with the effect of natural events. Detecting and reading house numbers from street views is a computer vision problem that falls under the category of natural scene text recognition. Convolutional neural network (CNN) model is one of the most commonly used deep learning (DL) methods in image analysis. In this study, firstly, CNN based DL method was applied to read characters from pictures that contain house numbers in their natural image. However, successful results could not be obtained, especially in cases where there are more than one house number in the image or when the depths are very variable. A new approach DDL (deep in deep learning) using two different CNN models was proposed to increase the accuracy of the DL method and also to reduce the data size created by natural images. The performance of the proposed DDL approach was compared with the performance of the DL approach using real data consisting of 17,618 images with 113 GB (gigabyte) size consisting of building images with GPS location information taken from 35 neighborhoods of Kayseri Metropolitan Municipality (KBB). Experimental results showed that the proposed DDL approach produced more accurate results and used less storage space than DL approach. [ABSTRACT FROM AUTHOR]

Hyperspectral images (HSI) are contiguous band images commonly used in remote sensing. Deep learning (DL) is an effective method used in HSI classification. Convolutional neural networks (CNN) are one of the DL methods used in HSI classification. It provides automated approaches that can learn abstract features of HSIs from spectral-spatial fields. The high dimensionality of the HSIs increases the computational complexity. Therefore, most of the developed CNN models perform dimensionality reduction as a preprocessing step. Another problem in HSI classification is that spectral-spatial features must be considered in order to obtain accurate results. Because, HSI classification performance is highly dependent on spectralspatial information. In this study, a new 3D CNN model is proposed for HSI classification. The proposed method provides an effective method to simultaneously extract spectral-spatial features in HSIs. The network uses the 3D hyperspectral cube at the input. Principal component analysis is used to eliminate the dimensional redundancy in the hyperspectral cube. Then, using neighborhood extraction, spectral-spatial features are extracted effectively. The proposed method has been tested with 4 datasets. The application results were compared with 7 different DL-based methods and it was seen that our 4CF-Net method showed better classification performance. [ABSTRACT FROM AUTHOR]

Shoulder implants must be replaced after a certain period of time. But determining the manufacturer or model of the implant during this change is often a mistake-prone and difficult process for medical experts. The aim of this study is to identify 4 different implant manufacturers from 597 shoulder implant X-ray images. For this purpose, both pre-trained CNN architectures (DenseNet201, DenseNet169, InceptionV3, NasNetLarge, VGG16, VGG19 and Resnet50) and cascade models that feed these architectures with YOLOv3 detection algorithm were created and the classification performances of these models were compared. The task of the YOLOv3 detection algorithm in cascade models is to detect the head area of the shoulder implants and give this area as an input to CNN architectures. In addition, traditional machine learning methods were combined with the ensemble learning method and their performance on the data set was revealed. The highest classification performance was achieved in the cascade DenseNet201 model with an accuracy rate of 84.76%. This rate is higher in the literature than in a different study using similar dataset. Classification accuracy of ensemble models is substantially lower than CNN models. Also, the classification accuracy of YOLO supported cascade models is higher than individual CNN models. That is, focusing on the head of the implant with the YOLOV3 detection algorithm has increased the classification accuracy. This method will inspire future studies in this field. [ABSTRACT FROM AUTHOR]

Although powder bed fusion joining (TYB) metal additive manufacturing is frequently preferred in the production of complex geometry parts today, real-time monitoring of part manufacturing processes is insufficient. Therefore, the machine control system remains largely open loop. While some metal additive manufacturing machines present the powder bed monitoring with images, it has not been found that they can automatically detect the defects that may occur in the powder bed layer and stimulate the control system. In the study, an exemplary machine learning-based approach is presented for on-site monitoring and defect detection of powder bed images, which can be a component of a real-time control system in any TYB metal additive manufacturing machine. Using the deep learning method, which is one of the subfields of machine learning, a classification was made to detect the defects that may occur in creating a layer of the process. Detection and classification of defects were carried out using the convolutional neural networks model. The data set for training and performance of the model was created with photographs of a three-dimensional sample structure manufactured on the EOS M290 machine. The best performance was obtained in the VGG16 model with 88.3% accuracy by performing transfer learning from VGG-16, Inception V3, and DenseNet pre-learning models. [ABSTRACT FROM AUTHOR]

In the literature, effective and common technology used for in door navigation has not been found. For the solution of the problem, recognition of the text images used indoors and determination of the current location according to the images seems to be the most feasible solution. In this study, deep learning methods, which are used successfully in many fields, were used to recognize signage images used in interior spaces. The created model extracts high-level features such as corners and edges in the images with the Convolutional Neural Network. Next, the sequencing properties of the letters are preserved with the Recurrent Neural Network-Based Bidirectional Long-Short-Term Memory. The Connectionist Temporal Classification approach is used to solve the repetitions and other problems in the sequence. Mistakes in words created with the characters generated by the trained model are improved using Levenshtein distance. The success of the model in the tests performed with the Synth90k data set was 96%, and the success in Turkish images with the same character set of the model trained with the Sytnh90kwas 93%. The results obtained demonstrate the success of the proposed approach. [ABSTRACT FROM AUTHOR]

The COVID-19 is a virus that spreads quickly with a high mortality rate. Rapid and accurate early diagnosis has a key role to reduce the mortality and to decrease the economic cost of this pandemic. For this purpose, diagnostic kits and diagnosis using medical imaging methods have been investigated. Among the medical imaging tools, diagnosis with the help of Computed Tomography and X-ray images is very important. Three different ResNet models (ResNet 50, ResNet 101, and ResNet 152) were investigated (a) to discriminate patients with COVID-19 from normal subjects, (b) to discriminate patients with COVID-19 from patients with Pneumonia, and (c) to discriminate patients with COVID-19, patients with Pneumonia, and normal subjects. ResNet 50 model gave the highest performances among these three models. As a result, we achieved the accuracy of 99.3% to discriminate COVID-19 and Normal, the accuracy of 99.2% to discriminate COVID-19 and Pneumonia, and the accuracy of 97.3% to discriminate COVID-19, Normal, and Pneumonia. In conclusion, the pre-trained ResNet 50 model has a big potential to detect the patients with COVID-19 quickly and accurately using chest X-Ray images only. We believe that this study will help to defeat the epidemic. [ABSTRACT FROM AUTHOR]

Covid-19 has been described as a pandemic by the World Health Organization. It has become an epidemic all over the world and has created a risk for people that may lead to death. To diagnose Covid-19, the diagnosis must be confirmed by RT-PCR test. The test takes a long time and false-negative results can be obtained. If the diagnosis of Covid-19 is made early and correct, the ratio of threats to life is reduced. Deep learning has been widely used in a variety of applications to solve a variety of complex problems that require extremely high accuracy and precision, especially in the medical field. In this study, the Covid-19 is diagnosed automatically using a proposed multi-channel CNN method. Patients and healthy individuals' Lung X-ray images datasets were obtained from three separate online databases. Simple recurrent networks (SRN) architecture was also applied for the same problem to compare the results and demonstrate the efficiency of the proposed method. It is to be noted that to reveal the performance, accuracy and efficiency of the study, accuracy and precision analysis and measurements of processing times for the applied methods were performed. With the proposed system Covid-19 is diagnosed in a short time without waiting for the PCR test and precautions are taken before the virus increases its effect on the body and the risk of individuals' life. Differently from the studies in the literature, the multi-channel CNN architecture with five convolution channels is proposed and the channel selection formulas are presented which are used for selecting the most distinctive feature filters among the results produced by these channels. [ABSTRACT FROM AUTHOR]

Bees are one of the oldest living species in the world, having a major impact on the development of living species. The continuity of plants at the bottom of the food chain is directly related to the pollination of bees. Bees are a global insurance because of this characteristic. For this reason, it is very important to check the health status of bees. Depending on the technology developed nowadays, it is possible to control the health status of bees remotely with real-time image processing applications. In the study, feature extraction methods, which are the strengths of deep learning, were executed from two different arms and aggressive changes in images were detected. In the classification process; Instead of Softmax classifier based on probability calculation, multi-layer feedback artificial neural network (MLFB-ANN) model has been used. The success of the designed system has also been compared with the Softmax classifier. As a result of experimental studies, 93.07% success rate can be achieved with Softmax classifier for six different bee diseases on the same data set, while 95.04% success rate has been obtained with the developed system. In this study, a hybrid method based on deep learning methods was proposed for the classification of bee diseases and successful results were obtained. [ABSTRACT FROM AUTHOR]

Sonar used to determine the size, distance, direction and other features of an object using sound waves; It is widely used in underwater mining and oil exploration, submarine mapping, tracking fish shoals and mine detection. Feature extraction, selection, selection of appropriate algorithms and optimization of hyperparameters that should be used for the identification and classification of sonar signals are scientific problems that have been studied for many years. In this study, using three different mathematical transformations as an innovative approach, numerical representation of data in a different format is proposed and the performance of deep learning methods in this particular problem is compared with classical machine learning and statistical pattern recognition algorithms. Mathematical transformations of Markov Transform Field, Gramian Angular Field and Repetition Plot are used to express the data in image format from time series type. By using the transformed data obtained, deep learning algorithms were trained and the performance of the produced models and the results obtained with classical algorithms were compared with the help of metrics. It has been determined that the proposed time series data to image transformation approaches eliminate the need for feature extraction in problem solving and give the best results determined in the literature so far. It is considered that the proposed new approach can be applied not only to time series based classification problems, but also in different research areas, and will provide significant contributions to increase the performance of machine learning algorithms with the proposed mathematical transformations for the representation of data. [ABSTRACT FROM AUTHOR]

Although the architectural design process is carried out with the collaboration of experts who are experienced in many different areas from the main preferences to the detailing stage, the major decisions such as plan organization, mass design etc. are taken by the architect. Computer Aided Design (CAD) programs are generally effective after the major decisions of the design are taken. For this reason, it is common for the main decisions, taken during the design process, to be changed during the analysis of the structural system. In order to prevent this, in the early stages of architectural design, earthquake system awareness and structural system design should be included as an design input; as, the failure of the structural system which did not considered well in the architectural design phase leads to unexpected revisions in the implementation project phase and thus leads to serious losses in both time and cost. The aim of this study is to create an Irregularity Control Assistant (IC Assitant) that can provide architects general information about the appropriateness of structural system decisions to earthquake regulations in the early stages of design process by using the deep learning and image processing methods. In this way, correct decisions will be made in the early stages of the design and unexpected revisions that may occur during the implementation project phase will be prevented. [ABSTRACT FROM AUTHOR]

Automatic Identification System (AIS) is an observation and analysis system that has become compulsory nowadays due to the risks of maritime transportation such as collision, fire, and spillage of hazardous or polluting substances. In the literature, we can see the applications of basic mathematical models, statistical models and machine learning algorithms using AIS data in order to detect these dangers in advance and to make controlled and safe travel of ships. In this study, AIS data have been evaluated as time series, and accuracy comparisons have been made by being developed different models with Autoregressive Integrated Moving Average, Multilayer Perceptron (MLP) and Deep Recurrent Neural Networks (DRNN) beside traditional route estimation model. In addition, feature selection techniques have been weighted in MLP and RDNN models, and new algorithms have been proposed with these improving. Relief, Pearson's Correlation, Gain Ratio and Information Gain (IG) methods were used to compare the accuracy of the route and collision estimations. In order to be used in these accuracy tests, AIS data related into certain times of Çanakkale Strait and Marmara Sea were used. The results showed that all the approaches were close and high accuracy due to the linear movement of the ships in Çanakkale Strait. On the other hand, it has been observed that the best approach in the Marmara Sea was the improved DRNN with IG due to its irregular structure. [ABSTRACT FROM AUTHOR]

Image processing applications in real-time systems have become a popular topic in recent years. Deep learning methods, one of the sub-branches of artificial intelligence, and image processing algorithms used in the field of object detection from images can be used together. In this way, applications are developed in many areas such as autonomous cars, autonomous unmanned aerial vehicles, assist robot technologies, assistant technologies for disabled and elderly individuals. This study aims to detect the tactile paving surfaces with deep learning methods in order to design an assistive technology system that can be used by visually impaired individuals, autonomous vehicles and robots. Contrary to traditional image processing algorithms, deep learning methods and image processing algorithms are used together in this study. The YOLO-V3 model, which is one of the best methods of object detection, is combined with the DenseNet model to create the YOLOV3-Dense model. YOLO-V2, YOLO-V3 and YOLOV3Dense models were trained on the Marmara Tactile Paving Surface (MDPY) dataset, which was created by the researchers and included 4580 images and their performances were compared with each other on the test dataset. It was observed that YOLOV3-Dense model is better than other models in detecting tactile paving surface with 89% F1-score, 92% mean average Precision(mAP) and 81% IoU values. [ABSTRACT FROM AUTHOR]

Generative adversarial training has been one of the most active research topics and many researchers have conducted their studies on Generative Adversarial Network (GAN) shortly after it is claimed to be one of the most promising research area of the last decade by pioneers of the deep learning community. On the other hand, the idea behind generators has also reemerged autoencoder models such as Variational Autoencoder (VAE). Therefore, autoencoder models have gained their popularity back. Some restrictions of GAN models such as lack of inference mechanism, GAN and VAE based hybrid models have proposed addressing image generation. With the effect of these notions and studies, we have also considered VAE and GAN hybrid models. For obtaining synthetic but at the same time high-resolution handwritten-looking images without any training, Compositional Pattern Producing Network (CPPN) is adapted from previous studies for combining with VAE and adversarial training. For improving generation capabilities, an objective from a previous VAE/GAN model is also adapted for our VAE/CPGAN hybrid model. For analyzing the proposed model performance, baseline models such as GAN, VAE and VAE/GAN are also evaluated for comparisons. In this paper. it is clearly seen the proposed model is capable of the generating realistic and scalable super resolution synthetic images on a common dataset. [ABSTRACT FROM AUTHOR]

Recently, detection and recognition of animal faces, body postures, behaviors, and physical movements is became an interdisciplinary field. Computer vision methods can contribute to determine behaviors of animals and predict the following behavior of animals. Moreover, these methods would contribute to domesticate animals. In this study, a deep learning based system is proposed for the detection and classification of dog's behaviour. In the study, firstly, a dataset is created by collecting videos containing the behavior of dogs which don't avoid contact with people. After the necessary analysis on the obtained videos, a customized data set consisting of more meaningful sections is developed by extracting determined behaviors in videos. It is recognized the behavior with the Faster R-CNN (Faster Regional-Convolutional Neural Networks) by selecting key frames from these customized video sections. In the last stage, the related behaviors in videos are followed by video tracker after the behavior of the dog is recognized. As a result of experimental studies, the behaviors of dog such as opening the mouth, sticking out the tongue, sniffing, rearing the ear, swinging the tail and playing are examined and accuracy rates 94.00%, 98.00%, 99.33%, 99.33%, 98.00% and 98.67% are obtained for these behaviors, respectively. With the results obtained in the study, it is seen that our proposed method based on key frame selection and determination of regions of interest are successful in recognition the behavior of dogs. [ABSTRACT FROM AUTHOR]

In this paper, a new hyperspectral image classification method based on 2-dimensional complex Gabor filtering and deep convolutional neural networks is proposed. Specifically, as a deep learning model, convolutional neural network is aimed to extract distinctive high-level features. Deep-learned and Gabor feature extraction methodologies are simultaneously performed on the input hyperspectral samples. Gabor features are calculated by implementing complex Gabor filtering only on the first three principal components of the hyperspectral image. The proposed hybrid model uses Gabor transform to obtain local image features, such as edges, corners and texture. The Gabor features of the images are calculated at multiple orientations and frequencies. Then, deep features and Gabor features are fused to obtain a more robust and discriminative feature vector. Hybrid feature vector is used as input to a softmax classifier for hyperspectral image classification. The parameters of the proposed deep learning architecture are optimized using a small training set. Thus, the over-fitting problem of the proposed convolutional neural network has been reduced to some extent. Experiments performed on two popular hyperspectral datasets show that the proposed method can achieve better classification performance than some conventional methods. Classification results demonstrates that the proposed hybrid model is an efficient method for feature extraction and classification of hyperspectral images. [ABSTRACT FROM AUTHOR]

Skin cancer is common and a serious disease, which can lead to death if not treated in time. Melanoma is the rarest and most dangerous type of skin cancer. It causes the most deaths. As in all diseases, early and correct detection of skin cancer are very important. Computer Aided Diagnosis systems can help physicians and patients make better decisions. Especially, machine learning and deep learning use effectively in Computer Aided Diagnosis systems. In this study, an automatic detection system for melanoma is suggested. To illustrate the advantage of the proposed CNN model C4Net, a comprehensive experimental study has been carried out. In addition, the proposed C4Net has been compared with not only the existing deep learning algorithms such as AlexNet, GoogLeNet, ResNet and VGGNet but also conventional machine learning algorithms such as Artificial neural networks, k-Nearest neighbor algorithm and Support vector machine. In experimental studies, C4Net, which is designed as deep neural network model for melanoma detection, has obtained more classification accuracy than other methods with 96.94%. [ABSTRACT FROM AUTHOR]

Human deaths caused by brain tumors are increasing nowadays. The brain tumor can grow very fast and can get twice of it is usual size. Therefore, physician have to analysis the Magnetic Resonance (MR) images quickly. This step is vital for the diagnosis of cancer, for treatment planning and for evaluation of the treatment outcome. If the patient who has the tumor in his brain is not treated correctly and quickly, the patient's chance of survival may decrease and resulted in death. In this article, we were developed a computer-assisted automated tumor detection system that can assist the physician in detecting and locating the tumor easily from brain MR images. The developed system is based on Regional based Convolutional Neural Networks (RCNN), which is one of the deep learning architectures. Besides RCNN is a structure that uses the architecture of the Convolutional Neural Networks (CNN), it can be considered as a structure in which the interested region is given as input in addition to the input images. In the proposed method, different RCNN architectures were designed and tested on Benchmark, Rembredant and Harvard datasets. The highest accuracy was obtained from the RCNN4 architecture on Benchmark data set is 99.10%. The highest average accuracy was calculated as 98.66% with RCNN4 architecture. Also, the success of the proposed method was compared with some of the methods exist in the literature. These comparisons showed that the proposed method is more successful and effective. [ABSTRACT FROM AUTHOR]

Human deaths caused by brain tumors are increasing nowadays. The brain tumor can grow very fast and can get twice of it is usual size. Therefore, physician have to analysis the Magnetic Resonance (MR) images quickly. This step is vital for the diagnosis of cancer, for treatment planning and for evaluation of the treatment outcome. If the patient who has the tumor in his brain is not treated correctly and quickly, the patient's chance of survival may decrease and resulted in death. In this article, we were developed a computer-assisted automated tumor detection system that can assist the physician in detecting and locating the tumor easily from brain MR images. The developed system is based on Regional based Convolutional Neural Networks (RCNN), which is one of the deep learning architectures. Besides RCNN is a structure that uses the architecture of the Convolutional Neural Networks (CNN), it can be considered as a structure in which the interested region is given as input in addition to the input images. In the proposed method, different RCNN architectures were designed and tested on Benchmark, Rembredant and Harvard datasets. The highest accuracy was obtained from the RCNN4 architecture on Benchmark data set is 99.10%. The highest average accuracy was calculated as 98.66% with RCNN4 architecture. Also, the success of the proposed method was compared with some of the methods exist in the literature. These comparisons showed that the proposed method is more successful and effective. [ABSTRACT FROM AUTHOR]