deep learning

“AppTree” is an intelligent platform to bring researchers, visitors, and all interested people closer to the oldest and most attractive botanical garden at the University of Tehran. AppTree can scan the QR-Barcode of each plant in person by smartphone or search various plants on the website and get all the useful knowledge about them. Also, the ability of AppTree is the recognition of different plants which don’t have labels. The plant recognition part is a machine learning module that can identify more than 100 different species of plants and give the user details about them. This novel platform is based on Android and Web-app and the identification of new plants type is done by machine learning approach. We utilized VGG19, a deep CNN, to classify images and to identify unlabeled plants. The classification accuracy, F1-score, recall, and precision were 98.25, 93.16%, 88.21%, and 94.85%, respectively, on the plant dataset of the University of Tehran. The proposed method was compared with other deep learning architectures such as AlexNet, AlexNetOWTBn, and GoogLeNet on the same dataset and obtained higher performance. Our AppTree platform has achieved considerable success and easily can be extended to use in other botanical gardens.

Review spam is an opinion written to promote or demote a product or brand on websites and other internet services by some users. Since it is not easy for humans to recognize these types of opinions, a model can be provided to detect them. In recent years, much research has been done to detect these types of reviews, and with the expansion of deep neural networks and the efficiency of these networks in various issues, in recent years, multiple types of deep neural networks have been used to identify spam reviews. This paper reviews the proposed deep learning methods for the problem of review spam detection. Challenges, evaluation criteria, and datasets in this area are also examined.

Since the outbreak of Covid19 virus to date, various methods have been introduced in order to diagnose the virus infection. One of the most reliable tests is assessing frontal Chest X-Ray(CXR) images. As the virus causes inflammation in the infected patient's lung, it is possible to diagnose whether one is infected or not using his/her CXR image. in contrast to other tests which mostly are based on the virus genome, this test is not time-consuming and it is reliable against new strains of the virus. However, this test requires a specialist to assess the CXR images. As the datasets of  Covid19 patient CXR images are increasing in number, it is possible to use machine learning techniques in order to assess CXR images automatically and even online. In this study, we used deep learning approaches and we fine-tuned the Alexent in order to automatically classify CXR images and label the whether "Covid" or "Normal". The data we used in this study include about 10,000 chest images, half of which are related to CXR images and the other half are related to patients with Covid19 infection. The model proved to be very reliable with 99.26% accuracy in diagnosis and 95% sensitivity and 99.7% specificity.

Falls are a serious concern among the elderly due to being a major cause of harm to their physical and mental health. Despite their potential for harm, they can be prevented with proper care and monitoring. As such, the motivation for this research is to implement an algorithmic solution to the problem of falls that leverages the benefits of Machine Learning to detect falls in the elderly. There are various studies on fall detection that works on one dataset: wearable, environmental, or vision. Such an approach is biased against low fall detection and has a high false alarm rate. According to the literature, using two datasets can result in high accuracy and lower false alarms. The purpose of this study is to contribute to the field of Machine Learning and Fall Detection by investigating the optimal ways to apply common machine and deep learning algorithms trained on multimodal fall data. In addition, it has proposed a multimodal approach by training two separate classifiers using both Machine and Deep Learning and combining them into an overall system using sensor fusion in the form of a majority voting approach. Each trained model outputs an array comprising three percentage numbers, the average of the numbers in the same class from both arrays is then computed, and the highest percentage is the classification result. The working system achieved results were 97% accurate, with the highest being achieved by the Convolutional Neural Network algorithm. These results were higher than other state-of-the-art research conducted in the field.

In recent scenario, women are suffering from breast cancer disease across the world. Mammography is one of the important methods to detect breast cancer early; that to reduce the cost and workload of radiologists. Medical image processing is a tremendous technique used to determine the disease in advance to reduce the risk factor. To predict the disease from 2-D mammography images for diagnosing and detecting based on advanced soft computing paradigm. Still, to get more accuracy in all coordinate axes, 3-D mammography imaging is used to capture depth information from all different angles. After the reconstruction of this process, a better quality of 3D mammography is obtained. It is useful for the experts to identify the disease in well advance. To improve the accuracy of disease findings, deep convolution neural networks (CNN) can be applied for automatic feature learning, and classifier building. This work also presents a comparison of the other state of art methods used in the last decades.

The aim of this study is to evaluate the performance of the pre-trained models and compare them with the probability percentage of prediction in terms of execution time. This study uses the COCO dataset to evaluate both pre-trained image recognition and object detection, models. The results revealed that Tiny-YoloV3 is considered the best method for real-time applications as it takes less time. Whereas ResNet 50 is required for those applications which require a high probability percentage of prediction, such as medical image classification. In general, the rate of probability varies from 75% to 90% for the large objects in ResNet 50. Whereas in Tiny-YoloV3, the rate varies from 35% to 80% for large objects, besides it extracts more objects, so the rise of execution time is sensible. Whereas small size and high percentage probability makes SqueezeNet suitable for portable applications, while reusing features makes DenseNet suitable for applications for object identification.

Accurate stock market prediction can assist in an efficient portfolio and risk management. However, accurately predicting stock price trends still is an elusive goal, not only because the stock market is affected by policies, market environment, and market sentiment, but also because stock price data is inherently complex, noisy, and nonlinear. Recently, the rapid development of deep learning can make the classifiers more robust, which can be used to solve nonlinear problems. This study proposes a hybrid framework using Long Short-Term Memory, Autoencoder, and Deep Neural Networks (LSTM-AE-DNNs). Specifically, LSTM-AE is responsible for extracting relevant features, and in order to predict price movement, the features are fed into two deep learning models based on a recurrent neural network (RNN) and multilayer perceptron (MLP). The dataset used for this is Dow Jones daily stock for 2008-2018, which was used in this article. Besides, to further assess the prediction performance of the proposed model, original stock features are fed to the single RNN and MLP models. The results showed that the proposed model gives the more accurate and best results compared to another. In particular, LSTM-AE+RNN shows a better performance than the LSTM-AE+MLP. In addition, hybrid models show better performance compared to a single DNN fed with the all-stock features directly.

Lung infection is the most dangerous sign of Covid 19. X-ray images are the most effective means of diagnosing this virus. In order to detect this disease, deep learning algorithms and machine vision are widely used by computer scientists. Convolutional neural networks (CNN), DenseNet121, Resnet50, and VGG16 were used in this study for the detection of Covid-19 in X-ray images. In the current study, 1341 chest radiographs from the COVID-19 dataset were used to detect COVID-19 including infected and Healthy classes using a modified pre-trained CNN (train and test accuracy of 99.75% and 99.63%, respectively). The DENSENET121 model has a training accuracy of 43.89% and a test accuracy of 57.89%, respectively. The train and test accuracy of ResNet-50 are, respectively, 89.43% and 90%. Additionally, the CNN model has test and train accuracy of 98.13% and 96.73%, respectively. The suggested model has COVID-19 detection accuracy that is at least 1% higher than all other models.

The Industrial Internet of Things (IIoT) is a potential platform for developing industry 4.0 and its related applications, especially in cyber-physical systems. Such a new trend in manufacturing sectors offers further potential to optimize operations, realize business models, and reduce costs. Such accomplish may also lead to complex and complicated tasks; hence, to deal with such issues, Reference Architecture Model Industry 4.0 (RAMI 4.0) is developed to structure Industry 4.0. In this paper, the standardized framework is considered RAMI 4.0 and its integration with an IIoT software named Software Platform Embedded Systems (SPES). Integrating Model-Based Engineering (MBE) with a framework requires using a deep learning model called Recurrent Neural Network (RNN). The RNN-MBE, which optimizes the entire process, is responsible for optimizing the process and reducing industry costs. The optimization problem has been fixed, and the MBE simulation has shown that using the proposed MBE is efficient.

Extracting sentiments from the English-Telugu code-mixed data can be challenging and is still a relatively new research area. Data obtained from the Twitter API has to be in English-Telugu code-mixed language. That data is free-form text, noisy, lexicon borrowings, code-mixed, phonetic typing and misspelling data. The initial step is language identification and sentiment class labels assigned to each tweet in the dataset. The second step is the data normalization task, and the final step is classification, which can be achieved using three different methods: lexicon, machine learning, and deep learning. In the lexicon-based approach, tokenize each tweet with its language tag. If the language tag is in Telugu, transliterate the roman script into native Telugu words. Words are verified with TeluguSentiWordNet, and the Telugu sentiments are extracted, and English SentiWordNets are used to extract sentiments from the English tokens. In this paper, the aspect-based sentiment analysis approach is suggested and used with normalized data. In addition, deep learning and machine learning techniques are applied to extract sentiment ratings, and the results are compared to prior work.

Due to the rapid advancements in computer technology, researchers are attracted to solving challenging problems in many different fields. The price of rainbow options is an interesting problem in financial fields and risk management. When there is no closed-form solution to some options, numerical methods must be used. Choosing a suitable numerical method involves the most appropriate combination of criteria for speed, accuracy, simplicity and generality. Monte Carlo simulation methods and traditional numerical methods have expensive repetitive computations and unrealistic assumptions on the model. Deep learning provides an effective and efficient method for options pricing. In this paper, the closed-form formula or Monte-Carlo simulation are used to generate data in European and Asian rainbow option prices for the deep learning model. The results confirm that the deep learning model can price the rainbow options more accurately with less computation time than Monte-Carlo simulation.

The widespread use of E-commerce websites has drastically increased the need for automatic recommendation systems with machine learning. In recent years, many ML-based recommenders and analysers have been built; however, their scope is limited to using a single filtering technique and processing with clustering-based predictions. This paper aims to provide a systematic year-wise survey and evolution of these existing recommenders and analysers in specific deep learning-based hybrid filtering categories using movie datasets. They are compared to others based on their problem analysis, learning factors, data sets, performance, and limitations. Most contributions are found with collaborative filtering using user or item similarity and deep learning for the IMDB datasets. In this direction, this paper introduces a new and efficient Hybrid Filtering based Recommendation System using Deep Learning (HFRS-DL), which includes multiple layers and stages to provide a better solution for generating recommendations.

Since 2010, deep learning has been further developed, and the concept of multi-modality has penetrated into all walks of life. However, it has not been fully researched and applied in college English teaching, so this study modeled and practiced the multimodal teaching method of college English under the deep learning mode and its application. The definitions of modality and medium are first introduced, and then the definition of multimodality in this study is clarified. Then the classification of multimodal transport is expounded. The random forest algorithm is chosen as the main algorithm of this research, and a dynamic multimodal model is established. After that, there was a collaboration with a university and sophomore students were selected for practice. After processing and analyzing the collected data, it was found that in the data sample of 268 students, the number of students who did not study independently accounted for 24%, which indicates that most college students lack interest in learning English. Preliminary tests were also conducted on students' English proficiency throughout the year, and the results showed that the students' English proficiency was at a pass level and the overall English proficiency was weak. Reassessment of students' English proficiency showed that the actual teaching effect of each English proficiency was greater than 85%, and the effectiveness of English teaching in the selected universities was significantly improved. The average score improved by 8 points, indicating that multimodal teaching is scientifically effective After a semester of multimodal teaching, the English teaching effectiveness of the university selected in this article has significantly improved. The research results indicate that the development of deep computer learning has introduced multimodal concepts into the teaching field, which is very suitable for assisting language learning based on its own advantages. After a semester of multimodal teaching, the English teaching effectiveness of the university selected in this article has significantly improved. The research results indicate that the development of deep computer learning has introduced multimodal concepts into the teaching field, which is very suitable for assisting language learning based on its own advantages.

Finding a brain tumor yourself by a human in this day and age by looking through a large quantity of magnetic-resonance-imaging (MRI) images is a procedure that is both exceedingly time consuming and prone to error. It may prevent the patient from receiving the appropriate medical therapy. Again, due to the large number of image datasets involved, completing this work may take a significant amount of time. Because of the striking visual similarity that exists between normal tissue and the cells that comprise brain tumors, the process of segmenting tumour regions can be a challenging endeavor. Therefore, it is absolutely necessary to have a system of automatic tumor detection that is extremely accurate. In this paper, we implement a system for automatically detecting and segmenting brain tumors in 2D MRI scans using a convolutional-neural-network (CNN), classical classifiers, and deep-learning (DL). In order to adequately train the algorithm, we have gathered a broad range of MRI pictures featuring a variety of tumour sizes, locations, forms, and image intensities. This research has been double-checked using the support-vector-machine (SVM) classifier and several different activation approaches (softmax, RMSProp, sigmoid). Since "Python" is a quick and efficient programming language, we use "TensorFlow" and "Keras" to develop our proposed solution. In the course of our work, CNN was able to achieve an accuracy of 99.83%, which is superior to the result that has been attained up until this point. Our CNN-based model will assist medical professionals in accurately detecting brain tumors in MRI scans, which will result in a significant rise in the rate at which patients are treated.

Heart and circulatory system diseases are often referred to as cardiovascular disease (CVD). The health and efficiency of the heart are crucial to human survival. CVD has become a primary cause of demise in recent years. According to data provided by the World-Health-Organization (WHO), CVD were conscientious for the deaths of 18.6M people in 2017. Biomedical care, healthcare, and disease prediction are just few of the fields making use of cutting-edge skills like machine learning (ML) and deep learning (DL). Utilizing the CVD dataset from the UCI Machine-Repository, this article aims to improve the accuracy of cardiac disease diagnosis. Improved precision and sensitivity in diagnosing heart disease by the use of an optimization algorithm is possible. Optimization is the process of evaluating a number of potential answers to a problem and selecting the best one. Support-Machine-Vector (SVM), K-Nearest-Neighbor (KNN), Naïve-Bayes (NB), Artificial-Neural-Network (ANN), Random-Forest (RF), and Gradient-Descent-Optimization (GDO) are just some of the ML strategies that have been utilized. Predicting Cardiovascular Disease with Intelligence, the best results may be obtained from the set of considered classification techniques, and this is where the GDO approach comes in. It has been evaluated and found to have an accuracy of 99.62 percent. The sensitivity and specificity were likewise measured at 99.65% and 98.54%, respectively. According to the findings, the proposed unique optimized algorithm has the potential to serve as a useful healthcare examination system for the timely prediction of CVD and for the study of such conditions.

Accurate traffic classification is important for various network activities such as accurate network management and proper resource utilization. Port-based approaches, deep packet inspection, and machine learning are widely used techniques for classifying and analyzing network traffic flows. Most classification methods are suitable for small-scale datasets and cannot achieve a high classification accuracy owing to their shallow learning structure and limited learning ability. The emergence of deep learning technology and software-driven networks has enabled the application of classification methods for processing large-scale data.In this study, a two-step classification method based on deep learning algorithms is presented, which can achieve high classification accuracy without manually selecting and extracting features. In the proposed method, an Autoencoder was used to extract features and remove unnecessary and redundant features. In the second step, the proposed method uses the features extracted by the autoencoder from a hybrid deep-learning model based on the CNN and LSTM algorithms to classify network traffic.To evaluate the proposed method, the results of the proposed two-stage hybrid method is compared with comparative algorithms including decision tree, Naïve Bayes, random forest. The proposed combined CNN+LSTM method obtains the best results by obtaining values of 0.997, 0.972, 0.959, and 0.964, respectively, for the evaluation criteria of, accuracy, precision, recall, and F1 score.The proposed method is a practical and operational method with high accuracy, which can be applied in the real world and used in the detection of security anomalies in networks using traffic classification and network data.

Sentiment analysis is the process of identifying and categorizing people’s emotions or opinions regarding various topics. The analysis of Twitter sentiment has become an increasingly popular topic in recent years. In this paper, we present several machine learning and a deep learning model to analysis sentiment of Persian political tweets. Our analysis was conducted using Bag of Words and ParsBERT for word representation. We applied Gaussian Naive Bayes, Gradient Boosting, Logistic Regression, Decision Trees, Random Forests, as well as a combination of CNN and LSTM to classify the polarities of tweets. The results of this study indicate that deep learning with ParsBERT embedding performs better than machine learning. The CNN-LSTM model had the highest classification accuracy with 89 percent on the first dataset and 71 percent on the second dataset. Due to the complexity of Persian, it was a difficult task to achieve this level of efficiency. The main objective of our research was to reduce the training time while maintaining the model's performance. As a result, several adjustments were made to the model architecture and parameters. In addition to achieving the objective, the performance was slightly improved as well.

With the advent of user-generated text information on the Internet, text sentiment analysis plays an essential role in online business transactions. The expression of feelings and opinions depends on the domains, which have different distributions. In addition, each of these domains or so-called product groups has its vocabulary and peculiarities that make analysis difficult. Therefore, different methods and approaches have been developed in this area. However, most of the analysis involved a single-domain and few studies on cross-domain mood classification using deep neural networks have been performed. The aim of this study was therefore to examine the accuracy and transferability of deep learning frameworks for the cross-domain sentiment analysis of customer ratings for different product groups as well as the cross-domain sentiment classification in five categories “very positive”, “positive”, “neutral”, “negative” and “very negative”. Labels were extracted and weighted using the Long Short-Term Memory (LSTM) Recurrent Neural Network. In this study, the RNN LSTM network was used to implement a deep transfer learning framework because of its significant results in sentiment analysis. In addition, two different methods of text representation, BOW and CBOW were used. Based on the results, using deep learning models and transferring weights from the source domain to the target domain can be effective in cross-domain sentiment analysis.