Contents of Volume 33 (2023)

1/2023 2/2023 3/2023 4/2023 5/2023


  • [1] Turk F., Akkur E., Erogul O., (Turkey)
    BI-RADS categories and breast lesions classification of mammographic images using artificial intelligence diagnostic models, pp. 413-432

      Full text

    Abstract: According to BI-RADS criteria, radiologists evaluate mammography images, and breast lesions are classified as malignant or benign. In this retrospective study, an evaluation was made on 264 mammogram images of 139 patients. First, data augmentation was applied, and then the total number of images was increased to 565. Two computer-aided models were then designed to classify breast lesions and BI-RADS categories. The first of these models is the support vector machine (SVM) based model, and the second is the convolutional neural network (CNN) based model. The SVM-based model could classify BI-RADS categories and malignant-benign discrimination with an accuracy rate of 86.42% and 92.59%, respectively. On the other hand, the CNN-based model showed 79.01% and 83.95% accuracy for BI-RADS categories and malignant benign discrimination, respectively. These results showed that a well-designed machine learning-based classification model can give better results than a deep learning model. Additionally, it can be used as a secondary system for radiologists to differentiate breast lesions and BI-RADS lesion categories.

  • [2] Tolic A., Boshkoska B.M., Skansi S. , (Slovenia, Croatia) ,
    Upgrading the JANET neural network by introducing a new storage buffer of working memory, pp. 433-459

      Full text

    Abstract: Recurrent neural networks (RNNs), along with long short-term memory networks (LSTMs), have been successfully used on a wide range of sequential data problems and have been entitled as extraordinarily powerful tools for learning and processing such data. However, the search for a new or derived architecture that would model very long-term dependencies is still an active area of research. In this paper, a relatively psychologically plausible architecture named event buffering JANET (EB-JANET) is proposed. The architecture is derived from the forgetgate- only version of the LSTM, which is also called just another network (JANET). The new architecture implements a new working memory mechanism that operates on information represented as dynamic events. The event buffer, as a container of events, is a reference to the state of the relevant pre-activation values on the basis of which historical candidate values were generated relative to the current timestep. The buffer is emptied as needed and depending on the context of information. The proposed architecture has achieved world-class results and it outperforms JANET on multiple benchmark datasets. Moreover, the new architecture is applicable to a wider class of problems and showed superior resilience when processing longer sequences, as opposed to JANET which experienced catastrophic failures on certain tasks.

  • [3] Aravind T., Suresh P., (India)
    Development of an efficient deep learning system for automatic prediction of power demand based on the forecasting of power distribution, pp. 461-479

      Full text

    Abstract: Electrical load prediction aids electrical producing and allocation firms in planning capacity and management to ensure that all customers get the energy they need on a consistent basis. Despite the fact that numerous prediction models have been created, none of them can be applied to all market trends. As a result, this article provides a practical technique for predicting customer power usage. To address the troubles of power utilization surveying, CRF-based energy utilization choosing strategy conditional random field based powered consumption prediction (CRF-PCP) is proposed. A convolutional brain organization (a technique in view of artificial intelligence) joined with a contingent irregular field is utilized to prepare and foresee the energy consumption (EC) of the districts. The training model’s features are extracted using a spatiotemporal texture map (STTM). Supervisory control and data acquisition (SCADA) is utilized to gather and keep up with information on the power utilization of local purchasers. The information given in the cloud is sent to the power circulation framework. Additionally, power utilization expectation utilizing a convolution neural network (CNN) with profound conditional random field (CRF) provides an outcome of 98.9% precision, which is far superior to prior research in the same area. The acquired result demonstrates that the employed machine learning methods are performing at their peak.

  • [4] Divya J., Chandrasekar A., (India)
    DRGNN – Dilated recurrent graph neural network framework incorporating spatial and temporal features signifying social relationships in IoT network based traffic prediction, pp. 481-499

      Full text

    Abstract: The intelligent transportation system seeks to reduce traffic and improve the driving experience. They give us a lot of data that we can use to improve services for both the public and transportation officials by feeding it into machine learning systems. Most importantly, Traffic environment refers to everything that might have an impact on how much traffic is moving down the road, including traffic signals, accidents, protests, and even road repairs that might result in a backup. A motorist or rider can make an informed choice if they have previous knowledge that is very close to approximate all the above and many more real-world circumstances that can affect traffic. Additionally, it aids in the development of driverless vehicles. Traffic data have been growing dramatically in recent decades, and we are moving toward big data concepts for transportation. The current approaches for predicting traffic flow use some traffic prediction models, however they are still inadequate to handle practical situations. We thus aimed to focus on the traffic flow forecast problem using the traffic data and prediction models. The proposed model called DRGNN, a dilated recurrent graph neural network framework aims to effectively analyze and predict the traffic pattern by considering the spatial (space) and temporal (time) aspects of the real-time traffic data considering social relationships between internet of vehicles which indeed produced accurate and valuable insights that could help in deploying the model in any suitable real-time traffic monitoring and prediction system.

  • [5] Contents volume 33 (2023), ... 501
  • [6] Authors index volume 33 (2023), ... 503


  • [1] Haroon M.S., Ali H.M., (Pakistan)
    Ensemble adversarial training based defense against adversarial attacks for machine learning-based intrusion detection system, pp. 317-336

      Full text

    Abstract: In this paper, a defence mechanism is proposed against adversarial attacks. The defence is based on an ensemble classifier that is adversarially trained. This is accomplished by generating adversarial attacks from four different attack methods, i.e., Jacobian-based saliency map attack (JSMA), projected gradient descent (PGD), momentum iterative method (MIM), and fast gradient signed method (FGSM). The adversarial examples are used to identify the robust machine-learning algorithms which eventually participate in the ensemble. The adversarial attacks are divided into seen and unseen attacks. To validate our work, the experiments are conducted using NSLKDD, UNSW-NB15 and CICIDS17 datasets. Grid search for the ensemble is used to optimise results. The parameter used for performance evaluations is accuracy, F1 score and AUC score. It is shown that an adversarially trained ensemble classifier produces better results.

  • [2] Vivek V., Hemalatha J., Latchoumi T.P., Mohan S. , (India)
    Towards the development of obstacle detection in railway tracks using thermal imaging, pp. 337-355

      Full text

    Abstract: To prevent collisions between trains and objects on the railway line, rugged trains require an intelligent rail protection system. To improve railway safety and reduce the number of accidents, studies are underway. Machine learning (ML) had progressed rapidly, creating new perspectives on the subject. A technique called speed up robust features (SURF) is proposed by researchers to collect regionally and globally relevant information. In addition, taking advantage of the Ohio State University (OSU) heat walker benchmarking dataset, the effectiveness of this technique was examined under various lighting scenarios. This technology could help in reducing train accident rates and financial costs. The findings of the proposed methodology are very specific, in addition to the ability to quickly identify items (obstacles) on the railway line, both of which contribute to rail security. The proposed faster region based convolutional neural network (FR-CNN) with 2D singular spectrum analysis (SSA) improves the performance analysis of an accuracy of 90.2%, recall 95.6% and precision 94.6% when compared with an existing system YOLOv2 and YOLOv5 with 2D SSA.

  • [3] Pozzobon E., Weiß N., Mottok J., Matoušek V., (Germany, CZ) ,
    An evolutionary fault injection settings search algorithm for attacks on safe and secure embedded systems, pp. 357-374

      Full text

    Abstract: In this paper, we present a novel method for exploiting vulnerabilities in secure embedded bootloaders, which are the foundation of trust for modern vehicle software systems, by using a genetic algorithm to successfully identify the correct parameters to perform an electromagnetic fault injection attack. Specifically, we demonstrate the feasibility of code execution attacks by leveraging a combination of software and hardware weaknesses in the secure software update process of electronic control units (ECUs), which is standardized across the automotive industry. Our method utilizes an automated approach, eliminating the need for static code analysis, and does not require any hardware modifications to the targeted systems. Through our research, we successfully demonstrated our attack on three distinct ECUs from different manufacturers used in current vehicles. Our results prove that the use of a genetic algorithm for finding the fault parameters reduces the number of attempts necessary for a successful fault to obtain arbitrary code execution via “wild jungle jumps” by approximately 100 times compared to a naive random search.

  • [4] Kekula F., Hrubeš P., (CZ)
    An empirical study of relationships between urban lighting indicators and night-time light radiance, pp. 375-396

      Full text

    Abstract: Night-time light (NTL) radiance has a great potential in analyses of dynamic changes in patterns of human activities, and socio-economic and demographic factors. However, most of those analyses are focused on factors at global scales such as the population size, gross domestic product, electric power consumption, fossil fuel carbon dioxide emission etc. In this study we investigate the relationships between three urban lighting indicators and monthly averaged NTL radiance obtained from NASA’s Black Marble monthly NTL composites for 4 study areas in the Czech Republic at local scale. The Pearson correlation analysis was used to identify a strength of the correlations between the indicators and radiance at near-nadir for two different snow conditions. The results from the correlation show that radiance has a strong positive correlation with the number of streetlighting points and their total nominal power, while for the average mast height there were observed moderate correlation coefficients. However, the areas with larger scales have higher correlation coefficients. Moreover, we found that the correlation coefficients are higher for snow-covered condition radiances. Generalized linear (GL) regression analysis was used to examine an association between the radiance and selected indicators. Owing to the excess zeros and overdispersion in the data, the zero-inflated regression performs better than the GL regression. Results from the regression analysis evince a statistically significant relationship between the radiance and selected indicators.

  • [5] Bělinová Z., Votruba Z., (CZ)
    Reflection on systemic aspects of consciousness, pp. 397-412

      Full text

    Abstract: Today’s quick development of artificial intelligence (AI) brings us to the questions that have until recently been the domain of philosophy or even sciencefiction. When can be a system considered an intelligent one? What is a consciousness and where it comes from? Can systems gain consciousness? It is necessary to have in mind, that although the development seems to be a revolutionary one, the progress is successive, today’s technologies did not emerge from thin air, they are firmly built on previous findings. As now some wild thoughts and theories where the AI development leads to have arisen, it is time to look back at the background theories and summarize, what do we know on the topics of intelligence, consciousness, where they come from and what are different viewpoints on these topics. This paper combines the findings from different areas and present overview of different attitudes on systems consciousness and emphasizes the role of systems sciences in helping the knowledge in this area.


  • [1] Baskar K., Vijayalakshmi P., Muthumanickam K., Arthi A., (India)
    A novel authentication and access scheduling scheme to improve the performance of WSN, pp. 205-224

      Full text

    Abstract: Wireless sensor network (WSN) is a kind of network specifically suitable for place where infrastructure and resources are playing a vital role. Moreover, nodes in a WSN are autonomous in nature. WSNs can be able to solve various real-time problems and issues like smart healthcare, smart office, smart energy, smart home, etc. As energy becomes one of the scarce supplies for this kind of network, attacks against authentication help to validate the legitimacy of sensor nodes become foremost important. Such attacks exhaust the power of nodes that are currently connected to a WSN, thereby reducing their lifetime. In this article, a zonal node authentication technique as well as optimal data access scheduling that renders data deliverance with improved quality of service and network lifetime is proposed. The results obtained from simulation for diverse WSN topologies accentuate the claim of our method over the existing solutions and demonstrate to be efficient in discovering legitimate sensor nodes with the optimal workload. Besides improved network lifetime, efficiency, and throughput, the proposed method also reinforces the security measures of the WSN by integrating node authentication.

  • [2] Poonkuzhali S., Shobana M., Jeyalakshmi J., (India)
    A deep transfer learning approach for IoT/IIoT cyber attack detection using telemetry data, pp. 225-244

      Full text

    Abstract: The rise of internet connectivity across the globe increases the count of IoT (internet of things)/IIoT (industrial internet of things) devices exponentially. The objects/devices which are connected to the internet are always prone to malicious attacks at various levels, such as physical, network, fog, and applications, which exist in the IoT architecture. Many researchers have addressed this issue and designed their own solutions based on machine and deep learning techniques. It is undeniable that deep learning outperforms machine learning (ML), but it necessitates a massive amount of datasets with appropriate labels. In this work, the deep transfer learning (TL) technique has been adapted for gated recurrent unit (GRU). Each model is trained using a dataset that belongs to one source IoT device (source domain), and this trained model is used to classify the malicious traffic in another dataset that belongs to some other IoT device (target domain). This approach is used for binary classification. These transfer learning models have been evaluated using an IoT/IIoT telemetry dataset called ToN IoT which comprises the sensor data generated from the seven different types of IoT devices. The highest accuracy achieved by IoT garage door was upto 99.76% as a source domain by fixing IoT thermostat as target domain. These models were also evaluated using some more metrics such as precision, recall, F1-measure, training time and testing time. By implementing transfer learning based GRU model, the accuracy of the model is improved from 69.20% to 99.76%. Moreover, to prove the efficiency of the proposed model, it is compared with state of art deep learning model and its results were analyzed in a detailed manner.

  • [3] Makhrus F., (Saudi Arabia)
    The effect of amplitude modification in S-shaped activation functions on neural network regression, pp. 245-269

      Full text

    Abstract: Time series forecasting using multilayer feed-forward neural networks (FNN) is potential to give high accuracy. Several factors influence the accuracy. One of them is the choice of activation functions (AFs). There are several AFs commonly used in FNN with their specific characteristics, such as bounded type AFs. They include sigmoid, softsign, arctan, and tanh. This paper investigates the effect of the amplitude in the bounded AFs on the FNNs’ accuracy. The theoretical investigations use simplified FNN models: linear equation and linear combination. The results show that the higher amplitudes give higher accuracy than typical amplitudes in softsign, arctan, and tanh AFs. However, in sigmoid AF, the amplitude changes do not influence the accuracy. These theoretical results are supported by experiments using the FNN model for time series prediction of 10 foreign exchanges from different continents compared to the US dollar. Based on the experiments, the optimum amplitude of the AFs should be high, that is greater or equal to 100 times of the maximum input values to the FNN, and the accuracy gains up to 3–10 times.

  • [4] Daqrouq K., Hazazi A., Alkhateeb A., Alharbey R.A., (Indonesia)
    Heart rate measurement using image recognition technology, pp. 271-290

      Full text

    Abstract: The measurement of heart rate (HR) has numerous applications in various fields, such as the internet of things, security, sports, and telemedicine. There are many methods for measuring pulse rates, and this research is based on a novel technique of measuring the heartbeat using image recognition technology. The innovations in the field of visual objects have made the detection process easy and quick, with high efficiency. Four step-based algorithms, including a computer, an external high-definition camera, and an open-source computer vision library, have been presented for measuring heart rate. The first step was the face detection (FD) algorithm, and the second was the area attention algorithm to determine the region of interest (ROI). The ROI signal analysis algorithm was used in the third step, using a fast Fourier transform (FFT) for frequency detection. The pulse measurement phase was the final step, and it was based on the strength of the color concentration in proportion to the time extracted from video clips. With the help of our recorded database of 50 participants based on different ages and skin colors, the process was carried out. The results of this study contributed to the development of an HR detection technique based on image recognition using the Python programming language. This is a very comfortable and effective method for measuring the human heart rate. This research article discussed various factors and obstacles that affect heart rate measurement. The results found that our system is highly competent in measuring heart rate.

  • [5] Uglickich E., Nagy I., (CZ)
    Using Poisson proximity-based weights for traffic flow state prediction, pp. 291-315

      Full text

    Abstract: The development of traffic state prediction algorithms embedded in intelligent transportation systems is of great importance for improving traffic conditions for drivers and pedestrians. Despite the large number of prediction methods, existing limitations still confirm the need to find a systematic solution and its adaptation to specific traffic data. This paper focuses on the relationship between traffic flow states in different urban locations, where these states are identified as clusters of traffic counts. Extending the recursive Bayesian mixture estimation theory to the Poisson mixtures, the paper uses the mixture pointers to construct the traffic state prediction model. Using the predictive model, the cluster at the target urban location is predicted based on the traffic counts measured in real time at the explanatory urban location. The main contributions of this study are: (i) recursive identification and prediction of the traffic state at each time instant, (ii) straightforward Poisson mixture initialization, and (iii) systematic theoretical background of the prediction approach. Results of testing the prediction algorithm on real traffic counts are presented.


  • [1] Krishna K.P.R., Thirumuru R. (India)
    Enhanced QOS energy-efficient routing algorithm using deep belief neural network in hybrid falcon-improved ACO nature-inspired optimization in wireless sensor networks, pp. 113-141

      Full text

    Abstract: Wireless sensor networks (WSNs) have recently acquired prominence in a variety of applications such as remote monitoring and tracking. Since it is virtually hard to recharge the nodes in their remote deployment, also, the transmission of data from nodes to the base station requires a significant amount of energy. Thus, our research proposes a routing protocol, namely hybrid falcon-improved ACO Nature-Inspired Optimization using a deep learning model to reduce energy consumption while increases the network lifetime. In the developed model, initially, the falcon optimization technique is utilized to locate the best possible cluster heads in the quickest possible time. Furthermore, to improve the quality of service in routing optimization a new improved ACO has been proposed in which linear flexible operator and the premier operator are used to increasing the iteration speed. Finally, the optimum route is obtained through DBNN based on predicted energy. As a result, our proposed model gives a lifetime as 121 s and energy consumption as 0.041 J at 500 rounds when compared to the baseline approaches. Therefore, our proposed approaches provides better routing and improves the QoS as well as the energy consumption which increases the longevity of mobile nodes.

  • [2] Yildirim S., Sevim C., Kalkat M., (Turkey)
    Vibration analyses of railway systems using proposed neural predictors, pp. 143-159

      Full text

    Abstract: Due to travelling on railway systems; there are many gaps and problems in cross areas. Therefore; it is necessary and very important to establish intelligent crossing systems in such areas. On the other hand, it is not possible for trains to stop or brake immediately against an obstacle due to their high speed and inertia. For this reason, it is necessary to work on the safety/warning of the other main factors and necessities (pedestrians and vehicles) in level crossings. This experimental investigation is carried out by using an experimental real-time train and crossing systems. The main vibration parameters are analysed by using neural networks. First, the dynamics of the train-rail system related to level crossings are examined, and the vibrations created by the train on rails are measured at different speeds. Then three types of proposed neural networks predictors, Levenberg-Marquardt backpropagation (LMBP), scaled conjugate gradient backpropagation (SCGB) and BFGS quasi-Newton backpropagation (BFGS) are used to predict the vibration of the train-rail system. From the results, it is seen that the proposed LMBP is more suitable for analysing and predicting the vibration of the train-rail system. It is clear that the speeds of the trains approaching the level crossing can be estimated from the vibration of the trains on the rails.

  • [3] Borna K., Ghanbari R., (Iran)
    A self-adaptive deep learning-based model to predict cloud workload, pp. 161-169

      Full text

    Abstract: Predicting cloud workload is a problematic issue for cloud providers. Recent research has led us to a significant improvement in workload prediction. Although self-adaptive systems have an imperative impact on lowering the number of cloud resources, those still have to be more accurate, detailed and accelerated. A new self-adaptive technique based on a deep learning model to optimize and decrease the use of cloud resources is proposed. It is also demonstrated how to prognosticate incoming workload and how to manage available resources. The PlanetLab dataset in this research is used. The obtained results have been compared to other relevant designs. According to these comparisons with the state-of-theart deep learning methods, our proposed model encompasses a better prediction efficiency and enhances productivity by 5%.

  • [4] Ahmed S.H., Raza M., Kazmi M., Mehdi S.S., Rehman I., Qazi S.A., (Pakistan)
    Towards the next generation intelligent transportation system: A vehicle detection and counting framework for undisciplined traffic conditions, pp. 171-189

      Full text

    Abstract: Modern development in deep learning and computer vision techniques, intelligent transportation system (ITS) has emerged as a useful tool for building a traffic infrastructure in smart cities. Previously, several computer vision techniques have been proposed for vehicle recognition, which were limited in handling undisciplined, dense and laneless traffic conditions. Moreover, these frameworks did not incorporate many of the local vehicle configurations common in South Asian countries such as Pakistan, Bangladesh, and India. Considering the limitations of previous frameworks, this paper presents efficient vehicle detection and counting model for undisciplined conditions including dense and laneless traffic, occulusion cases and diverse range of local vehicles. A dataset of more than 2400 images of vehicles has been collected comprising of six new categories of local vehicles, and considering undisciplined traffic conditions to ensure robustness in vehicle detection and counting system. Transfer learning based technique has been used, using faster R-CNN model with Inception V2 as underlying architecture. The experimental results show a precision of 86.14% in terms of mAP. The work finds its application in South Asian contexts as more smart cities are formed in this region. The proposed framework will enable traffic monitoring with higher reliability, accuracy and granularity, contributing in having next-generation ITS.

  • [5] Abu-Qasmieh I.F., Masad I.S., Alquran H., Alawneh K.Z., (Jordan)
    Generation of synthetic FLAIR MRI image from real CT image for accurate synovial fluid segmentation in human knee image, pp. 191-203

      Full text

    Abstract: Synthetic MRI FLAIR images of an abdominal 3D multimodality phantom and in vivo human knee have been generated from real CT images using predefined mapping functions of CT mean and standard deviation with the corresponding proton density PD, T1 and T2 that were previously generated from spin-echo sequence. First, the validity of generating synthetic MR images from different sequences were tested and the same PD, T1 and T2 maps that were generated from the real CT image have been used in the simulation of MRI inversion-recovery (IR) sequence. The similarity results between the real and synthetic IR sequence images, using different inversion times TI, showed a very good agreement. After confirming the feasibility of generating synthetic IR images from the PD, T1 and T2-maps, that were originally obtained from spin-echo sequence using the phantom, the simulation of a knee image has been generated from the corresponding knee CT real image using the steady-state transverse magnetization formula of the inversion-recovery sequence. The simulated FLAIR IR sequence MR image are generated using proper TI for nulling the signal from the synovial fluid, where the image complement is used as a mask for segmenting the corresponding tissue region in the real CT image.


  • [1] Yumoto M., Hagiwara M. (Japan)
    Selective classification considering time series characteristics for spiking neural networks , pp. 49-66

      Full text

    Abstract: In this paper, we propose new methods for estimating the relative reliability of prediction and rejection methods for selective classification for spiking neural networks (SNNs). We also optimize and improve the efficiency of the RC curve, which represents the relationship between risk and coverage in selective classification. Efficiency here means greater coverage for risk and less risk for coverage in the RC curve. We use the model internal representation when time series data is input to SNN, rank the prediction results that are the output, and reject them at an arbitrary rate. We propose multiple methods based on the characteristics of datasets and the architecture of models. Multiple methods, such as a simple method with discrete coverage and a method with continuous and flexible coverage, yielded results that exceeded the performance of the existing method, softmax response.

  • [2] Akkur E., Türk F., Erogul O. (Turkey)
    Breast cancer classification using a novel hybrid feature selection approach, pp. 67-83

      Full text

    Abstract: Many women around the world die due to breast cancer. If breast cancer is treated in the early phase, mortality rates may significantly be reduced. Quite a number of approaches have been proposed to help in the early detection of breast cancer. A novel hybrid feature selection model is suggested in this study. This novel hybrid model aims to build an efficient feature selection method and successfully classify breast lesions. A combination of relief and binary Harris hawk optimization (BHHO) hybrid model is used for feature selection. Then, k-nearest neighbor (k-NN), support vector machine (SVM), logistic regression (LR) and naive Bayes (NB) methods are preferred for the classification task. The suggested hybrid model is tested by three different breast cancer datasets which are Wisconsin diagnostic breast cancer dataset (WDBC), Wisconsin breast cancer dataset (WBCD) and mammographic breast cancer dataset (MBCD). According to the experimental results, the relief and BHHO hybrid model improves the performance of all classification algorithms in all three datasets. For WDBC, relief-BHO-SVM model shows the highest classification rates with an of accuracy of 98.77%, precision of 97.17%, recall of 99.52%, F1-score of 98.33%, specificity of 99.72% and balanced accuracy of 99.62%. For WBCD, relief-BHO-SVM model achieves of accuracy of 99.28%, precision of 98.76%, recall of 99.17%, F1-score of 98.96%, specificity of 99.56% and balanced accuracy of 99.36%. Relief-BHO-SVM model performs the best with an accuracy of 97.44%, precision of 97.41%, recall of 98.26%, F1-score of 97.84%, specificity of 97.47% and balanced accuracy of 97.86% for MBCD. Furthermore, the relief-BHO-SVM model has achieved better results than other known approaches. Compared with recent studies on breast cancer classification, the suggested hybrid method has achieved quite good results.

  • [3] Sohaib M., Tehseen S. (Pakistan)
    Forgery detection of low quality deepfake videos, pp. 85-99

      Full text

    Abstract: The rapid growth of online media over different social media platforms or over the internet along with many benefits have some negative effects as well. Deep learning has many positive applications like medical, animations and cybersecurity etc. But over the past few years, it is observed that it is been used for negative aspects as well such as defaming, black-mailing and creating privacy concerns for the general public. Deepfake is common terminology used for facial forgery of a person in media like images or videos.The advancement in the forgery creation area have challenged the researchers to create and develop advance forgery detection systems capable to detect facial forgeries. Proposed forgery detection system works on the CNN-LSTM model in which we first extracted faces from the frames using MTCNN then performed spatial feature extraction using pretrained Xception network and then used LSTM for temporal feature extraction. At the end classification is performed to predict the video as real or fake. The system is capable to detect low quality videos. The current system has shown good accuracy results for detecting real or fake videos on the Google deepfake AI dataset.

  • [4] Zhang X., Zhao N., Lv Q., Ma Z., Qin Q., Gan W., Bai J., Gan L. (China)
    Garbage classification based on a cascade neural network, pp. 101-112

      Full text

    Abstract: Most existing methods of garbage classification utilize transfer learning to acquire acceptable performance. They focus on some smaller categories. For example, the number of the dataset is small or the number of categories is few. However, they are hardly implemented on small devices, such as a smart phone or a Raspberry Pi, because of the huge number of parameters. Moreover, those approaches have insufficient generalization capability. Based on the aforementioned reasons, a promising cascade approach is proposed. It has better performance than transfer learning in classifying garbage in a large scale. In addition, it requires less parameters and training time. So it is more suitable to a potential application, such as deployment on a small device. Several commonly used backbones of convolutional neural networks are investigated in this study. Two different tasks, that is, the target domain being the same as the source domain and the former being different from the latter, are conducted besides. Results indicate with ResNet101 as the backbone, our algorithm outperforms other existing approaches. The innovation is that, as far as we know, this study is the first work combining a pre-trained convolutional neural network as a feature extractor with extreme learning machine to classify garbage. Furthermore, the training time and the number of trainable parameters is significantly shorter and less, respectively.


  • [1] Guo H., Tao X., Li X. (China)
    Water quality image classification for aquaculture using deep transfer learning , pp. 1-18

      Full text

    Abstract: With the development of high-density and intensive aquaculture production and the increasing frequency of water quality changes in aquaculture water bodies, the number of pollution sources in aquaculture ponds is also increasing. As the water quality of aquaculture ponds is a crucial factor affecting the production and quality of pond aquaculture products, water quality assessment and management are more important than in the past. Water quality analysis is a crucial way to evaluate the water quality of fish farming water bodies. Traditional water quality analysis is usually obtained by practitioners through experience and visual observation. There is an observability deviation caused by subjectivity. Deep transfer learning-based water quality monitoring system is easier to deploy and can avoid unnecessary duplication of efforts to save costs for aquaculture industry. This paper uses the transfer learning model of artificial intelligence to analyze the water color image automatically. 5203 water quality images are collected to create a water quality image dataset, which contains five classes based on water color. Based on the dataset, a deep transfer learning-based classification model is proposed to identify water quality images. The experimental results show that the deep learning model based on transfer learning achieves 99% accuracy and has excellent performance.

  • [2] Hlubuček A. (CZ)
    Integration of railway infrastructure topological description elements from the microL2 to the macroN0,L0 level of detail, pp. 19-34

      Full text

    Abstract: The paper presents the method of integration, which is supposed to be applied to the structure of the railway infrastructure topological description system expressed at the level of detail designated as microL2 in order to transform it into the structure expressed at the level of detail designated as macroN0,L0 . The microL2 level is the level of detail at which individual tracks in the structural sense and turnout branches are recognized, while the macroN0,L0 level is the level of individual operational points and line sections. The proposed integration algorithm takes into account both the parameter values of the individual elements appearing at the reference level of detail microL2 and their topological interconnectedness. Based on these aspects, these elements are integrated into the elements of the derived level of detail macroN0,L0 that can be described by the transformed parameter values. The relations between the respective elements are also transformed accordingly. While describing the transformation algorithm, the terminology and principles of the UIC RailTopoModel are used.

  • [3] Xu Z.Z., Zhang W.J. (China)
    3D CNN hand pose estimation with end-to-end hierarchical model and physical constraints from depth images, pp. 35-42

      Full text

    Abstract: Previous studies are mainly focused on the works that depth image is treated as flat image, and then depth data tends to be mapped as gray values during the convolution processing and features extraction. To address this issue, an approach of 3D CNN hand pose estimation with end-to-end hierarchical model and physical constraints is proposed. After reconstruction of 3D space structure of hand from depth image, 3D model is converted into voxel grid for further hand pose estimation by 3D CNN. The 3D CNN method makes improvements by embedding end-to-end hierarchical model and constraints algorithm into the networks, resulting to train at fast convergence rate and avoid unrealistic hand pose. According to the experimental results, it reaches 87.98% of mean accuracy and 8.82 mm of mean absolute error (MAE) for all 21 joints within 24 ms at the inference time, which consistently outperforms several well-known gesture recognition algorithms.