Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558927
Ilkin Huseynli, Songül Varlı
Deepfakes allow users to manipulate the identity of a person in a video or an image. Improvements on GAN-based techniques also generate more realistic and hard to detect fake faces. This threatens individuals and decreases trust in social media platforms. In this work, our goal is to report eight different models’ learning ability on, by far, the largest fake face dataset - DFDC. The models’ generalization ability was tested on the DFDC test set and Celeb-DF-v2 dataset. Effect of the various cut-out like augmentations to the learning was also reported.
{"title":"Analyzing Deep Learning Models’ Generalization Ability Under Different Augmentations on Deepfake Datasets","authors":"Ilkin Huseynli, Songül Varlı","doi":"10.1109/UBMK52708.2021.9558927","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558927","url":null,"abstract":"Deepfakes allow users to manipulate the identity of a person in a video or an image. Improvements on GAN-based techniques also generate more realistic and hard to detect fake faces. This threatens individuals and decreases trust in social media platforms. In this work, our goal is to report eight different models’ learning ability on, by far, the largest fake face dataset - DFDC. The models’ generalization ability was tested on the DFDC test set and Celeb-DF-v2 dataset. Effect of the various cut-out like augmentations to the learning was also reported.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"38 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131953531","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558929
Ali Pişirgen, Serhat Peker
Online appointment is now a growing trend that leads developers to create appointment booking systems for various industries such as health, education, tourism, transportation, production, and beauty. While different conceptual unified modelling language (UML) models exist for each industry, this causes time and cost consumption for business and leads inefficient use of resources. This study, therefore, intends to provide a UML based conceptual model for appointment booking system that enable system analyst and developers to gain advantage with regards to system development activities. Three different UML diagrams are used to demonstrate the users, the relationship between users and system, the exchange of commends. With this study, using the proposed generic model acting like a bridge between developers and coding, appointment booking application can be easily developed.
{"title":"A UML-Based Conceptual Model for Appointment Booking Systems","authors":"Ali Pişirgen, Serhat Peker","doi":"10.1109/UBMK52708.2021.9558929","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558929","url":null,"abstract":"Online appointment is now a growing trend that leads developers to create appointment booking systems for various industries such as health, education, tourism, transportation, production, and beauty. While different conceptual unified modelling language (UML) models exist for each industry, this causes time and cost consumption for business and leads inefficient use of resources. This study, therefore, intends to provide a UML based conceptual model for appointment booking system that enable system analyst and developers to gain advantage with regards to system development activities. Three different UML diagrams are used to demonstrate the users, the relationship between users and system, the exchange of commends. With this study, using the proposed generic model acting like a bridge between developers and coding, appointment booking application can be easily developed.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"127 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134157133","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558926
I. Aliyev, J. Mack, Nirmal Kumbhare, A. Akoglu, H. F. Ugurdag
This paper proposes a new hardware scheduler. As heterogeneous computing becomes prevalent, mapping applications on to multiple processing elements (PEs) proves to be non-trivial. Heterogeneous Earliest Finish Time (HEFT) algorithm is an already existing scheduler that aims to minimize the total execution time of an application. The paradigm of HEFT is such that it accepts an acyclic task graph as input at run-time and assigns/schedules the precompiled atomic tasks to PEs. HEFT stands out among many such schedulers not only in terms of producing shorter schedules but also in terms of its own short execution time. However, in real-time applications, the lower the latency, the better it is. To the best of our knowledge, this work is the only work that implements HEFT in hardware (on FPGA) further lowering its latency from milliseconds to as much as less than a microsecond. Porting HEFT to hardware has been challenging as data dependencies limit the amount of parallelism. Design of an efficient memory access pattern as well as an “incremental sorter” were key enablers in reducing the latency of the hardware implementation. We also integrated our FPGA-HEFT into an ARM-based SoC and validated its functionality using a realistic workload.
{"title":"FPGA-based Minimal Latency HEFT Scheduler for Heterogeneous Computing","authors":"I. Aliyev, J. Mack, Nirmal Kumbhare, A. Akoglu, H. F. Ugurdag","doi":"10.1109/UBMK52708.2021.9558926","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558926","url":null,"abstract":"This paper proposes a new hardware scheduler. As heterogeneous computing becomes prevalent, mapping applications on to multiple processing elements (PEs) proves to be non-trivial. Heterogeneous Earliest Finish Time (HEFT) algorithm is an already existing scheduler that aims to minimize the total execution time of an application. The paradigm of HEFT is such that it accepts an acyclic task graph as input at run-time and assigns/schedules the precompiled atomic tasks to PEs. HEFT stands out among many such schedulers not only in terms of producing shorter schedules but also in terms of its own short execution time. However, in real-time applications, the lower the latency, the better it is. To the best of our knowledge, this work is the only work that implements HEFT in hardware (on FPGA) further lowering its latency from milliseconds to as much as less than a microsecond. Porting HEFT to hardware has been challenging as data dependencies limit the amount of parallelism. Design of an efficient memory access pattern as well as an “incremental sorter” were key enablers in reducing the latency of the hardware implementation. We also integrated our FPGA-HEFT into an ARM-based SoC and validated its functionality using a realistic workload.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134183426","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558919
Furkan Artunc, I. Oksuz
Endoscopy imaging is a clinical procedure for the early detection of numerous cancers as well as for therapeutic procedures and minimally invasive surgery. Using endoscopic examination data to detect diseases is very helpful for medical doctors and speeds up the diagnosis. Because of the very narrow area, captured frames during endoscopic examination include a variety of artefacts. Artefacts degrade diagnostic image quality, which in turn makes disease diagnosis difficult for both clinicians and computer aided disease detection algorithms. Therefore, it is very crucial to find and eliminate those artefacts from medical images. In this paper, a detection system which utilizes ensemble of deep learning models and data augmentation is proposed. A fast and accurate object detection model which is YOLOv5 (improved version of YOLOv4) is selected as a base model. The 3 separate models are trained with segregated and augmented data; then, the models are combined to make an ensemble. The EndoCV2020 dataset is utilized to benchmark the ensemble model. The model achieves state-of-the-art performance with 49.6 mAP. The final mAP is calculated averaging several APs for different IoU thresholds (starting from 0.25 IoU to 0.75 Iou with step size 0.05).
{"title":"An Ensemble Approach for Automatic Artefact Detection on Gastroendoscopy Images","authors":"Furkan Artunc, I. Oksuz","doi":"10.1109/UBMK52708.2021.9558919","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558919","url":null,"abstract":"Endoscopy imaging is a clinical procedure for the early detection of numerous cancers as well as for therapeutic procedures and minimally invasive surgery. Using endoscopic examination data to detect diseases is very helpful for medical doctors and speeds up the diagnosis. Because of the very narrow area, captured frames during endoscopic examination include a variety of artefacts. Artefacts degrade diagnostic image quality, which in turn makes disease diagnosis difficult for both clinicians and computer aided disease detection algorithms. Therefore, it is very crucial to find and eliminate those artefacts from medical images. In this paper, a detection system which utilizes ensemble of deep learning models and data augmentation is proposed. A fast and accurate object detection model which is YOLOv5 (improved version of YOLOv4) is selected as a base model. The 3 separate models are trained with segregated and augmented data; then, the models are combined to make an ensemble. The EndoCV2020 dataset is utilized to benchmark the ensemble model. The model achieves state-of-the-art performance with 49.6 mAP. The final mAP is calculated averaging several APs for different IoU thresholds (starting from 0.25 IoU to 0.75 Iou with step size 0.05).","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"128 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130925919","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558915
E. Bulus
Today, it is desired to make various inferences from images quickly with artificial intelligence methods. One of the most important reasons for this is the increase in social media environments. On the other hand, it is desired to evaluate the images with artificial intelligence methods for security purposes. In this study, it was investigated how much the determination of the gender of a person from a face photograph can be done with existing methods. For this purpose, two of the widely used Convolutional Neural Network (CNN) methods were selected. The selected methods are caffemodel and vggl6 model. The accuracy of both methods was tested for the prepared male and female face images.
{"title":"Gender Determination from Pictures with CNN Models","authors":"E. Bulus","doi":"10.1109/UBMK52708.2021.9558915","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558915","url":null,"abstract":"Today, it is desired to make various inferences from images quickly with artificial intelligence methods. One of the most important reasons for this is the increase in social media environments. On the other hand, it is desired to evaluate the images with artificial intelligence methods for security purposes. In this study, it was investigated how much the determination of the gender of a person from a face photograph can be done with existing methods. For this purpose, two of the widely used Convolutional Neural Network (CNN) methods were selected. The selected methods are caffemodel and vggl6 model. The accuracy of both methods was tested for the prepared male and female face images.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"28 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132506415","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9559015
Tuğçe Koçak, M. Aydın, Berna Kiraz
Examination, monitoring and analysis of structural changes in the blood vessels of the brain enable the observation of brain functions. Therefore, the segmentation of the entire cerebral vascular network (including the capillaries) is of great importance in terms of the relevant specialist’s opinion on the diagnosis and treatment of a disease. When performed manuall, segmentation of the vascular network of the brain is a long time-consuming and fault-tolerant process. The automatic segmentation of the brain microvascular structure with machine learning approaches eliminates the need for specialists, and provides a method for perfroming cerebral vessel segmentation in a short time. This study provides the empirical comparision of three different deep neural network models including autoencoder, U-Net and ResNet+U-Net for the vascular network segmentation of brain vessels. The experiments are conducted on vesseINN dataset, which is a volumetric cerebrovascular system dataset obtained by two-photon microscopy. The models are evaluated based on accuracy, f1-score, recall, and precision metrics. During the training phase, U-Net and ResNet+Unet achieve 98% accuracy. Auto-encoder, on the other hand, yields 95% accuracy. In the test phase, it is observed that U-Net and ResNet+U-Net models give better results than the autoencoder model, according to the results obtained with 97% accuracy for U-Net and ResNet+Unet networks and 95% accuracy for autoencoder.
{"title":"Empirical Comparison of Deep Neural Networks for Brain Vessel Segmentation","authors":"Tuğçe Koçak, M. Aydın, Berna Kiraz","doi":"10.1109/UBMK52708.2021.9559015","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9559015","url":null,"abstract":"Examination, monitoring and analysis of structural changes in the blood vessels of the brain enable the observation of brain functions. Therefore, the segmentation of the entire cerebral vascular network (including the capillaries) is of great importance in terms of the relevant specialist’s opinion on the diagnosis and treatment of a disease. When performed manuall, segmentation of the vascular network of the brain is a long time-consuming and fault-tolerant process. The automatic segmentation of the brain microvascular structure with machine learning approaches eliminates the need for specialists, and provides a method for perfroming cerebral vessel segmentation in a short time. This study provides the empirical comparision of three different deep neural network models including autoencoder, U-Net and ResNet+U-Net for the vascular network segmentation of brain vessels. The experiments are conducted on vesseINN dataset, which is a volumetric cerebrovascular system dataset obtained by two-photon microscopy. The models are evaluated based on accuracy, f1-score, recall, and precision metrics. During the training phase, U-Net and ResNet+Unet achieve 98% accuracy. Auto-encoder, on the other hand, yields 95% accuracy. In the test phase, it is observed that U-Net and ResNet+U-Net models give better results than the autoencoder model, according to the results obtained with 97% accuracy for U-Net and ResNet+Unet networks and 95% accuracy for autoencoder.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"40 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132719121","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9559034
M. B. Terzi
In this study, a robust anomaly detection technique for ECG signals is developed using deep gated recurrent neural networks (GRNN) with Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM) unit. By training deep GRU and LSTM networks on normal ECG data acquired from healthy subjects, a robust prediction model that learns to predict future time steps of ECG time series is developed. The prediction errors are modeled using Multivariate Gaussian Distribution and the estimations of optimum parameters were performed via Maximum Likelihood Estimation (MLE) method. By using probability distributions of prediction errors and optimum threshold values, the classification of normal and abnormal time series is performed. The results of the study show that deep LSTM networks with stacked recurrent hidden layers can learn higher-level temporal features in ECG time series without prior knowledge of the data and can robustly model normal time series behaviors. The performance results of the proposed deep learning and Gaussian-based statistical anomaly detection technique over the European ST-T database show that the technique provides the reliable diagnosis of cardiovascular diseases by performing the robust detection of anomalies in ECG time series.
{"title":"Anomaly Detection with Deep Long Short Term Memory Networks","authors":"M. B. Terzi","doi":"10.1109/UBMK52708.2021.9559034","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9559034","url":null,"abstract":"In this study, a robust anomaly detection technique for ECG signals is developed using deep gated recurrent neural networks (GRNN) with Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM) unit. By training deep GRU and LSTM networks on normal ECG data acquired from healthy subjects, a robust prediction model that learns to predict future time steps of ECG time series is developed. The prediction errors are modeled using Multivariate Gaussian Distribution and the estimations of optimum parameters were performed via Maximum Likelihood Estimation (MLE) method. By using probability distributions of prediction errors and optimum threshold values, the classification of normal and abnormal time series is performed. The results of the study show that deep LSTM networks with stacked recurrent hidden layers can learn higher-level temporal features in ECG time series without prior knowledge of the data and can robustly model normal time series behaviors. The performance results of the proposed deep learning and Gaussian-based statistical anomaly detection technique over the European ST-T database show that the technique provides the reliable diagnosis of cardiovascular diseases by performing the robust detection of anomalies in ECG time series.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"30 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114959492","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558984
R. Arslan
Feature selection as a dimension reduction technique aims to select the subset containing less features by removing unrelated redundant or noisy features. While feature selection generally provides a better recognition performance, it also brings significant gains in calculation cost. In this study, the effects of using the most up-to-date feature selection methods on Android malware detection are shown. In order to observe this effect, test sets in 90 different combinations were prepared and comprehensive experiments were carried out objectively. As a result of the tests, a 4% increase in classification performance was achieved with the recursive feature selection method(RFE), while the gain in calculation cost was 39.39% in the chi2 method. Feature selection in application security analysis in the Android both contributed to the success of classification and reduced the time needed for classification. With this study, it has been shown the feature selection methods are an improvement that can affect the results of studies on Android security.
{"title":"Comparison of Feature Selection Methods in Security Analysis of Android","authors":"R. Arslan","doi":"10.1109/UBMK52708.2021.9558984","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558984","url":null,"abstract":"Feature selection as a dimension reduction technique aims to select the subset containing less features by removing unrelated redundant or noisy features. While feature selection generally provides a better recognition performance, it also brings significant gains in calculation cost. In this study, the effects of using the most up-to-date feature selection methods on Android malware detection are shown. In order to observe this effect, test sets in 90 different combinations were prepared and comprehensive experiments were carried out objectively. As a result of the tests, a 4% increase in classification performance was achieved with the recursive feature selection method(RFE), while the gain in calculation cost was 39.39% in the chi2 method. Feature selection in application security analysis in the Android both contributed to the success of classification and reduced the time needed for classification. With this study, it has been shown the feature selection methods are an improvement that can affect the results of studies on Android security.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116041313","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Digital advertising agencies and advertisers place billions of ads on search network every day. Managing these ads brings a lot of workload. One of the biggest problems in the growing digital advertising industry is bid optimization. The selection of the target audience, the randomness of user inquiries, the determination of ads by the auction system are the main factors that complicate the optimization problem. Reinforcement learning algorithms have become popular with their structures that provide solutions to complex problems in the field of advertising optimization in recent years. In this study, we determined the device, age, city and gender information of the target audience that will maximize the conversion rate of the campaign by using the most basic of reinforcement learning algorithms which is epsilon greedy.
{"title":"Demographic Targeting With Epsilon-greedy Exploration in Digital Advertising","authors":"Basak Esin Köktürk Güzel, Bora Mocan, Büsra Arslan, Gokce Polat, Tarık Kavuşan","doi":"10.1109/UBMK52708.2021.9558951","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558951","url":null,"abstract":"Digital advertising agencies and advertisers place billions of ads on search network every day. Managing these ads brings a lot of workload. One of the biggest problems in the growing digital advertising industry is bid optimization. The selection of the target audience, the randomness of user inquiries, the determination of ads by the auction system are the main factors that complicate the optimization problem. Reinforcement learning algorithms have become popular with their structures that provide solutions to complex problems in the field of advertising optimization in recent years. In this study, we determined the device, age, city and gender information of the target audience that will maximize the conversion rate of the campaign by using the most basic of reinforcement learning algorithms which is epsilon greedy.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"198 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123560105","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-09-15DOI: 10.1109/UBMK52708.2021.9558907
Mehmet Bilge Han Tas, Muhammed Coskun Irmak, Sedat Turan, A. Hasiloglu
The study was carried out in order to enable systems with weak processing power and motion to detect objects using cloud services. In addition, the dataset is expanded by continuous labeling to create big data. In the study, it is aimed to detect objects using cloud-based deep learning methods with an unmanned aerial vehicle (UAV). In the study, training processes were carried out with Google Colaboratory, a cloud service provider. The training processes are a YOLO-based system, and a convolutional neural network was created by revising the parameters in line with the needs. The convolutional neural network model provides communication between neurons in the convolutional layers by bringing the image data to the desired pixel ranges. Unlabeled pictures are included in the training by being tagged. In this way, it is possible to continuously enlarge the data pool. Since the microcomputers used in UAVs are insufficient for these processes, a cloud-based training model has been created. As a result of the study, cloud-based deep learning models work as desired. It is possible to show the accuracy of the model with the low losses seen in the loss functions and the mAP value. Graphic cards with high processing power are needed to provide training. It is essential to use powerful graphics cards when working on image data. Cost reduced by using cloud services. The training was accelerated and high-rate object detections were made. YOLOv5x was used in the study. It is preferred because of its fast training and high frame rate. Recall 80% Precision 93% mAP 82.6% values were taken.
{"title":"Real-Time Puddle Detection Using Convolutional Neural Networks with Unmanned Aerial Vehicles","authors":"Mehmet Bilge Han Tas, Muhammed Coskun Irmak, Sedat Turan, A. Hasiloglu","doi":"10.1109/UBMK52708.2021.9558907","DOIUrl":"https://doi.org/10.1109/UBMK52708.2021.9558907","url":null,"abstract":"The study was carried out in order to enable systems with weak processing power and motion to detect objects using cloud services. In addition, the dataset is expanded by continuous labeling to create big data. In the study, it is aimed to detect objects using cloud-based deep learning methods with an unmanned aerial vehicle (UAV). In the study, training processes were carried out with Google Colaboratory, a cloud service provider. The training processes are a YOLO-based system, and a convolutional neural network was created by revising the parameters in line with the needs. The convolutional neural network model provides communication between neurons in the convolutional layers by bringing the image data to the desired pixel ranges. Unlabeled pictures are included in the training by being tagged. In this way, it is possible to continuously enlarge the data pool. Since the microcomputers used in UAVs are insufficient for these processes, a cloud-based training model has been created. As a result of the study, cloud-based deep learning models work as desired. It is possible to show the accuracy of the model with the low losses seen in the loss functions and the mAP value. Graphic cards with high processing power are needed to provide training. It is essential to use powerful graphics cards when working on image data. Cost reduced by using cloud services. The training was accelerated and high-rate object detections were made. YOLOv5x was used in the study. It is preferred because of its fast training and high frame rate. Recall 80% Precision 93% mAP 82.6% values were taken.","PeriodicalId":106516,"journal":{"name":"2021 6th International Conference on Computer Science and Engineering (UBMK)","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128473318","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：