Network RNN Research Articles

Retraction Note: Effective offline handwritten text recognition model based on a sequence-to-sequence approach with CNN–RNN networks

Retraction Note: Effective offline handwritten text recognition model based on a sequence-to-sequence approach with CNN–RNN networks

Crop Classification with Attention Based BI-LSTM and Temporal Convolution Neural Network Combination for Remote Sensing Breizhcrop Time Series Data

In the modern era, remote sensing data has become increasingly useful for determining land use and coverage requirements. Remote sensing data can be used for a variety of purposes, including the classification of crops. It is possible to aggregate remote sensing data for a specific area over time in order to obtain a more complete picture based on the time series of this data. One example of these types of data is the Breizhcrop dataset, which was collected using satellite images acquired by Sentinel 2 over a period of time. This study aims to investigate a neural network based on attention mechanisms using the BI-LSTM layer in conjunction with Temporal-CNN for the classification of crops. The aim of the research is to find a model for corps classification in image-based time series. In line with this goal, in addition to finding features over time, the presented model also needs to produce high-accuracy features at each time step to increase classification. Utilizing the designed neural network, we seek to find local features with the attention mechanism and general features with a second layer. This neural network was validated on the BreizhCrop dataset and we conclude that it performs better than alternative approaches. The proposed method has been compared with Temporal CNN, Star RNN, and Vanilla LSTM networks and it has obtained better results than the mentioned neural networks. Taking advantage of these local and global features that extract with developed model obtained a high accuracy rate of 82%.

Self-Enhanced Attention for Image Captioning

Image captioning, which involves automatically generating textual descriptions based on the content of images, has garnered increasing attention from researchers. Recently, Transformers have emerged as the preferred choice for the language model in image captioning models. Transformers leverage self-attention mechanisms to address gradient accumulation issues and eliminate the risk of gradient explosion commonly associated with RNN networks. However, a challenge arises when the input features of the self-attention mechanism belong to different categories, as it may result in ineffective highlighting of important features. To address this issue, our paper proposes a novel attention mechanism called Self-Enhanced Attention (SEA), which replaces the self-attention mechanism in the decoder part of the Transformer model. In our proposed SEA, after generating the attention weight matrix, it further adjusts the matrix based on its own distribution to effectively highlight important features. To evaluate the effectiveness of SEA, we conducted experiments on the COCO dataset, comparing the results with different visual models and training strategies. The experimental results demonstrate that when using SEA, the CIDEr score is significantly higher compared to the scores obtained without using SEA. This indicates the successful addressing of the challenge of effectively highlighting important features with our proposed mechanism.

Tabular text embedding for Vietnamese text-based person search

Vietnamese text-based person search is still a challenging problem with the limited dataset of Vietnamese descriptions. The current popular approach to this problem is Deep Neural Networks (DNNs), and recently, transformer networks have been more favored because of their outperformance over CNN and RNN networks for both vision and natural language processing tasks. However, DNN, or transformer networks, require a large amount of training data and computing time for efficient learning of visual and textual features. This brings a burden for implementing Vietnamese text-based person search by DNN, or transformer networks. Towards building a Vietnamese text-based person search system on a scarce resource dataset of Vietnamese descriptive sentences with low computing cost, in this work, we propose to apply the transformer-based architecture named TabTransformer for contextual embedding of the noun phrases chunked from the Vietnamese descriptive sentences. This is the first time the TabTransformer network has been deployed together with CNN and RNN architectures for Vietnamese text-based person search. The experimental results on a limited dataset of 3000VnPersonSearch show the better recognition accuracy of the proposed method compared to the baseline method by about 7.5% at Rank 1. In addition, the computing time of our method is more effective than the baseline method.

Optimizing RNNs for EMG Signal Classification: A Novel Strategy Using Grey Wolf Optimization.

Accurate classification of electromyographic (EMG) signals is vital in biomedical applications. This study evaluates different architectures of recurrent neural networks for the classification of EMG signals associated with five movements of the right upper extremity. A Butterworth filter was implemented for signal preprocessing, followed by segmentation into 250 ms windows, with an overlap of 190 ms. The resulting dataset was divided into training, validation, and testing subsets. The Grey Wolf Optimization algorithm was applied to the gated recurrent unit (GRU), long short-term memory (LSTM) architectures, and bidirectional recurrent neural networks. In parallel, a performance comparison with support vector machines (SVMs) was performed. The results obtained in the first experimental phase revealed that all the RNN networks evaluated reached a 100% accuracy, standing above the 93% achieved by the SVM. Regarding classification speed, LSTM ranked as the fastest architecture, recording a time of 0.12 ms, followed by GRU with 0.134 ms. Bidirectional recurrent neural networks showed a response time of 0.2 ms, while SVM had the longest time at 2.7 ms. In the second experimental phase, a slight decrease in the accuracy of the RNN models was observed, standing at 98.46% for LSTM, 96.38% for GRU, and 97.63% for the bidirectional network. The findings of this study highlight the effectiveness and speed of recurrent neural networks in the EMG signal classification task.

Forecasting greenhouse air and soil temperatures: A multi-step time series approach employing attention-based LSTM network

Greenhouses stand as key infrastructural components in contemporary agriculture, facilitating the perennial availability of vegetables. Harnessing accurate, real-time environmental data, particularly air and soil temperature, is pivotal to promoting the facility’s efficiency in pest and disease alarming and vegetable production. Nevertheless, despite consistent long-term temperature trends, short-term fluctuations prove to be considerably nonlinear, time-lagged, and multivariate-coupled due to the unique microclimate of the greenhouse. Addressing this, we propose a novel short-term multi-step prediction model that assimilates a range of environmental data including air temperature, humidity, soil temperature, and soil moisture content at diverse greenhouse heights. This model utilizes an Attention-LSTM based time series approach to accurately predict multi-step short-term temperature variations within the greenhouse. By employing a modest amount of historical data (approximately 48 h), the model provides future temperature forecasts within a range of 30 to 480 min with high accuracy. When benchmarked against renowned models such as RNN, GRU, and LSTM networks, our proposed models display impressive performance. It achieved R2 values of 0.93, 0.94, 0.95, and 0.86 for 1, 4, 8, and 16-step predictions, respectively, in the domain of air temperature forecasting. The corresponding MSE values stood at 0.42, 0.32, 0.27, and 0.81. In parallel, for soil temperature forecasts, our model recorded R2 values of 0.96, 0.94, 0.92, and 0.89 for 1, 4, 8, and 16-step predictions, respectively, alongside MSE values of 0.36, 0.48, 0.73, and 0.99 for the equivalent steps. To conclude, these results reveal the exceptional predictive aptitude of our model for both air and soil temperature forecasts within greenhouse environments, validating its standing as an effective tool for the optimization of vegetable farming practices. Consequently, our model demonstrates its ability to yield precise predictions of environmental variables within greenhouses, offering crucial early warning mechanisms for the enhancement of vegetable cultivation processes.

Fusion model with attention mechanism for carbon-neutral sports competitions

IntroductionIn sports competitions, using energy-saving and emission-reduction measures is an important means to achieve the carbon neutrality goal.MethodsIn this paper, we propose an attention mechanism-based convolutional neural network (CNN) combined with the gated recurrent unit (GRU) for carbon neutral energy saving and emission reduction prediction model in sports competitions, CNN is a feedforward neural network whose input is a two-dimensional matrix. The main feature of CNN is that it can handle multi-channel input data, and the use of GRU can make the model structure simple and largely reduce The use of GRU can make the model structure simple, which largely reduces the hardware computational power and time cost and also better solves the long dependency problem in RNN networks. CNN-GRU extracts the data for features and then optimized by the attention mechanism.ResultsThe model collects real-time carbon emissions data from sports events, including game times, lighting usage, air conditioning emissions and other carbon emissions data, and uses deep learning algorithms to predict and compare carbon emissions from sports competition.DiscussionIn identifying energy saving and emission reduction measures conducive to the realization of the goal of carbon neutral sports events, the model has a certain reference value for realizing energy saving and emission reduction in sports competitions under carbon neutrality goals.

Optimization of Deep Neural Networks Using DAPP (DNN Acceleration Using Ping-Pong) Approach

Abstract: Deep Neural Networks (DNNs) are one of the leading classification algorithms. Deep Learning has achieved remarkable milestones such as Google-Net and Alpha-Go. They have shown promising results in pattern recognition for images and text, language translation, sound recognition, and many more. DNN has been widely accepted and also employed for pattern matching and image recognition. All these applications are possible as these networks emulate functioning of human brain and hence name “Neural” Network.However, to provide competent results, these millions of neurons in neural networks needs to be trained. For which billions of operations are to be carried out. Training of these many neurons and with these many operations is a time-consuming affair. Hence choice of network and its parameter play an important role both in providing accurate trained network and time taken for training. If the network is deep and has plethora of neurons, the time taken is considerably high as training works sequentially on batches of dataset using Sequential Back-propagation Algorithm.To accelerate the training there are many hardware solutions like use of GPU, FPGA and ASICs. However because of popularity of DNN there is increase in demand in mobile and IoT platform devices. These are resource constrained devices, where power and size of these device, restricts usage and implementation of deep NN (Neural Network).Simulation of DAPP is done on MNIST and CIFAR-10 datasets using System-C. Additionally, this technique has been adapted for multi-core architectures. The design shows a reduction in time by 38% for 3 layers of CNN and 92% for 10 layers of CNN, while maintaining the accuracy of networks. This generic methodology has been implemented for Vanilla RNN and LSTM networks. An improvement of 38% for Vanilla RNN and 40% for LSTM has been demonstrated by this methodology.

금속탐지를 위한 CNN 및 RNN 네트워크 개발에 관한 연구

This paper is a study on the efficiency of the filtering method of signal processing and the metal detection method using deep learning for data obtained from multiple MI sensors. The MI sensor is a principle that detects changes in magnetic field and is a passive sensor that detects metal objects. However, when detecting a metal object, the amount of change in the magnetic field caused by the metal is small, so there is a limit to the detectable distance. In order to effectively detect and analyze this, a method using deep learning was applied. In addition, the performance of the deep learning model was compared and analyzed using the filtering method of signal processing. In this paper, the detection performance of CNN and RNN networks was compared and analyzed from the data extracted from the self-impedance sensor. The RNN model showed higher performance than the CNN model. However, in the shallow stage, the CNN model showed higher performance than the RNN model.

Multi-Label Attribute Selection of Arrhythmia for Electrocardiogram Signals with Fusion Learning.

There are three primary challenges in the automatic diagnosis of arrhythmias by electrocardiogram (ECG): the significant variation among individual patients, the multiple pathologies in the ECG signal and the high cost in annotating clinical ECG with the corresponding labels. Traditional ECG processing approaches rely heavily on prior knowledge, such as those from feature extraction and waveform analysis. The preprocessing for prior knowledge incurs computational overhead. Furthermore, standard deep learning methods do not fully consider the dynamic temporal, spatial and multi-labeling characteristics of ECG data. In clinical ECG waveforms, it is common to see multi-labeling in which a patient is labeled with multiple classes of arrhythmias. However, multiclass approaches in current research mainly solve the multi-label machine learning problem, ignoring the correlation between diseases, resulting in information loss. In this paper, an arrhythmia detection and classification scheme called multi-label fusion deep learning is proposed. The objective is to build a unified system with automatic feature learning which supports effective multi-label classification. First, a multi-label ECG-based feature selection method is combined with a matrix decomposition and sparse learning theory. The optimal feature subset is selected as a preprocessing algorithm for ECG data. A multi-label classifier is then constructed by fusing CNN and RNN networks to fully exploit the interactions and features of the time and space dimensions. The experimental result demonstrates that the proposed method can achieve a state-of-the-art performance compared to other algorithms in multi-label database experiments.

A Rapid Combined Model for Automatic Generating Web UI Codes

Encoder-Decoder network is usually applied to image caption to automatically generate descriptive text for a picture. Web user interface (Web UI) is a special type of image and is usually described by HTML (hypertext marked language). Consequently, it becomes possible to use the encoder-decoder network to generate the corresponding code from a screenshot of Web UI. The basic structure of the decoder is RNN, LSTM, GRU, or other recurrent neural networks. However, this kind of decoder needs a long training time, so it increases the time complexity of training and prediction. The HTML language is a typically structured language to describe the Web UI, but it is hard to express the timing characteristics of the word sequence and the complex context. To resolve these problems efficiently, a rapid combined model RCM (rapid combined model) is designed in this paper. The basic structure of the RCM is an encoder-decoder network. The word embedding matrix and visual model are included in the encoder. The word embedding matrix uses fully connected units. Compared with LSTM, the accuracy of the word embedding matrix is basically unchanged, but the training and prediction speed have been significantly improved. In the visual model, the pretrained InceptionV3 network is used to generate the image vector, which not only improves the quality of the recognition of the Web UI interface image but also reduces the training time of the RCM significantly. In the decoder, the word embedding vector and the image vector are integrated together and input into the prediction model for word prediction.

LSTM Encoder-Decoder Dropout Model in Software Reliability Prediction

Numerous methods have been introduced to predict the reliability of software. In general, these methods can be divided into two main categories, namely parametric (e.g. software reliability growth models) and non-parametric (e.g. neural networks). Both approaches have been successfully implemented in software testing applications over the past four decades. Since most software reliability prediction data are available in the form of time series, deep recurrent network models (e.g. RNN, LSTM, NARX, and LSTM Encoder-Decoder networks) are considered as powerful tools to be employed in reliability-related problems. However, the problem of overfitting is a major concern when using deep neural networks for software reliability applications. To address this issue, we propose the use of dropout; therefore, this study utilizes a deep learning model based on LSTM Encoder-Decoder Dropout to predict the number of faults in software and assess software reliability. Experimental results show that the proposed model has better prediction performance compared with other RNN-based models.

A sequence models-based real-time multi-person action recognition method with monocular vision

In intelligent video surveillance under complex scenes, it is vital to identify the current actions of multi-target human bodies accurately and in real time. In this paper, a real-time multi-person action recognition method with monocular vision is proposed based on sequence models. Firstly, the key points of multi-target human body skeleton in the video are extracted by using the OpenPose algorithm. Then, the human action features are constructed, including limb direction vector and the skeleton height-width ratio. The multi-target human bodies tracking is then achieved by using the tracking algorithm. Next, the tracking results are matched with the action features, and the action recognition model is constructed, which includes the spatial branch based on Deep neural networks and the temporal branch based on Bi-directional RNN and Bi-directional long short-term memory networks. After pre-training, the model can be used to recognize the human body action from action features, and a recognition stabilizer is designed to minimize false alarms. Finally, extensive evaluations on the JHMDB dataset validate the effectiveness and the superiority of the proposed approach.

A Study on the Performance Comparison of Neural Network for Metal Detection Based on MI Sensor

This paper is a study on the efficiency of the filtering method of signal processing and the metal detection method using deep learning for data obtained from multiple MI sensors. The MI sensor is a principle that detects changes in magnetic field and is a passive sensor that detects metal objects. However, when detecting a metal object, the amount of change in the magnetic field caused by the metal is small, so there is a limit to the detectable distance. In order to effectively detect and analyze this, a method using deep learning was applied. In addition, the performance of the deep learning model was compared and analyzed using the filtering method of signal processing. In this paper, the detection performance of CNN and RNN networks was compared and analyzed from the data extracted from the self-impedance sensor. The RNN model showed higher performance than the CNN model. However, in the shallow stage, the CNN model showed higher performance than the RNN model.

G2Basy: A framework to improve the RNN language model and ease overfitting problem

Recurrent neural networks are efficient ways of training language models, and various RNN networks have been proposed to improve performance. However, with the increase of network scales, the overfitting problem becomes more urgent. In this paper, we propose a framework—G2Basy—to speed up the training process and ease the overfitting problem. Instead of using predefined hyperparameters, we devise a gradient increasing and decreasing technique that changes the parameters training batch size and input dropout simultaneously by a user-defined step size. Together with a pretrained word embedding initialization procedure and the introduction of different optimizers at different learning rates, our framework speeds up the training process dramatically and improves performance compared with a benchmark model of the same scale. For the word embedding initialization, we propose the concept of “artificial features” to describe the characteristics of the obtained word embeddings. We experiment on two of the most often used corpora—the Penn Treebank and WikiText-2 datasets—and both outperform the benchmark results and show potential towards further improvement. Furthermore, our framework shows better results with the larger and more complicated WikiText-2 corpus than with the Penn Treebank. Compared with other state-of-the-art results, we achieve comparable results with network scales hundreds of times smaller and within fewer training epochs.

Air quality prediction method based on improved wavelet denoising and LSTM

in order to optimize the traditional wavelet de-noising algorithm threshold function in continuity and noise reduction effect, an improved soft hard threshold compromise method and deep learning model are proposed to analyze the prediction performance of different deep learning models and improved wavelet algorithm. RNN, GRU and LSTM neural networks are constructed to predict the original data, the traditional wavelet denoising data and the improved wavelet denoising data. The air quality data of Chengdu were used for simulation experiment. The results of quantitative analysis showed that: for AQI, PM2.5 and O3_ Oh denoising will cause the loss of useful signals. The improved wavelet threshold denoising for PM10, SO2, Co, NO2 data can improve the prediction effect of the model. The results show that LSTM model has good applicability in air quality prediction, the absolute percent error is 5.867%, and the mean square error is 4.870. The prediction performance of LSTM model is about 10%, the absolute percent error is 5.176%, and the mean square error is 5.314. The traditional threshold denoising will affect the prediction results of the model for the actual data due to the loss of high-frequency signal and pseudo Gibbs phenomenon. The results show that the improved soft hard threshold compromise method can better deal with air quality data, and the LSTM model based on improved wavelet denoising has strong applicability for air quality prediction.

Deep Neural Network--based Machine Translation System Combination

Deep neural networks (DNNs) have provably enhanced the state-of-the-art natural language process (NLP) with their capability of feature learning and representation. As one of the more challenging NLP tasks, neural machine translation (NMT) becomes a new approach to machine translation and generates much more fluent results compared to statistical machine translation (SMT). However, SMT is usually better than NMT in translation adequacy and word coverage. It is therefore a promising direction to combine the advantages of both NMT and SMT. In this article, we propose a deep neural network--based system combination framework leveraging both minimum Bayes-risk decoding and multi-source NMT, which take as input the N-best outputs of NMT and SMT systems and produce the final translation. In particular, we apply the proposed model to both RNN and self-attention networks with different segmentation granularity. We verify our approach empirically through a series of experiments on resource-rich Chinese⇒English and low-resource English⇒Vietnamese translation tasks. Experimental results demonstrate the effectiveness and universality of our proposed approach, which significantly outperforms the conventional system combination methods and the best individual system output.

Lightweight versus obfuscation-resilient malware detection in android applications

By increasing growth of mobile applications, providing their security has become significant. Among mobile operating systems, Android is the most popular one, and hence, it has drawn more attention from malware programmers. One of the main challenges in designing a malware detection mechanism is handling obfuscation, where malware programmers try to change malware codes, such that they cannot be detected by malware detectors, while they keep their functionalities. In this paper, we propose an obfuscation-resilient method, called ORDroid, which can detect mutated and transformed malwares. We have used RNN and NLP neural networks for achieving this purpose. Our assumption is that the model is run on a server, before the application is published for end users. Users may get an application from different sources, and hence, it is necessary to design methods that can run on end users’ mobile phones. The challenge that should be considered when designing such methods is the limitation of computation and energy resources on a mobile phone. In the second part of this paper, we propose a lightweight malware detection method, called LightDroid. The main idea of this method is to select a minimal number of features from AndroidManifest file, along with a number of picture-based features from Dalvik executable file in a way that the accuracy of the resulting model is close to the state-of-the-art methods, while its complexity is as low as possible. We have fully implemented our proposed methods, as well as some of the state-of-the-art methods, including Drebin and RevealDroid. The results show that LightDroid is the most lightweight one, with 97.49% accuracy on the test data. Evaluation of ORDroid shows that, considering the overall accuracy of both test and transformed data, our model is the best comparing to the most related methods with the accuracy of 98.07% on the normal and 93.00% on the transformed data.

The assessment of 3D model representation for retrieval with CNN-RNN networks

In this paper, we propose a novel method for assessing 3D model representation via CNN and RNN networks. First, a visual tool developed with OpenGL is utilized to extract virtual views of each 3D m...

A novel ANN-based harmonic extraction method tested with ESN, RNN and MLP in shunt active power filters

With the wide use of power conversion devices - 'nonlinear loads' - many harmonic currents are being injected into the power grid. Shunt Active Power Filters SAPF are the power electronic equipment to compensate the harmonic currents caused by nonlinear loads. As the foundation of the harmonics recognition and compensation, harmonic extraction is the key technology in SAPF. Artificial Neural Networks ANN method has the features of parallel computation and satisfactory results for distorted source voltages over traditional extraction methods. This paper proposes a new harmonic extraction method based on ANN. To test the feasibility of different types of neural networks in this application, this paper compares the performances of three types of ANN: Echo State Networks ESN, Recurrent Neural Networks RNN and Multilayer Perceptron Networks MLP.