Search

Retinal images are a key element for ophthalmologists in diagnosing a variety of eye illnesses. The retina is vulnerable to microvascular changes as a result of many retinal diseases and a variety of research have been done on early diagnosis of medical images to take proper treatment on time. This paper designs an automated deep learning-based non-invasive framework to diagnose multiple eye diseases using colour fundus images. A multi-class eye disease RFMiD dataset was used to develop an efficient diagnostic framework. Multi-class fundus images were extracted from a multi-label dataset and then various augmentation techniques were applied to make the framework robust in real-time. Images were processed according to the network for low computational demand. A multi-layer neural network EyeDeep-Net has been developed to train and test images for diagnosis of various eye problems in which the keystone convolutional neural network extracts relevant features from the input colour fundus image dataset and then processed features were used to make predictive diagnostic decisions. The strength of the EyeDeep-Net is evaluated using multiple statistical parameters and the performance of the proposed model is found to be significantly superior to multiple baseline state-of-the-art models. A comprehensive comparison of the proposed methodology to the most recent methods proves its efficacy in terms of classification and disease identification through digital fundus images. [ABSTRACT FROM AUTHOR]

Human emotions play a vital role in developing social interaction as well as interpersonal relationships. It is because humans express and convey their thoughts and feelings visually rather than using verbal communication. Recently, facial expression recognition is greatly adopted in the emerging human–computer interaction and human–robot interaction technologies. Although humans can identify facial emotions virtually, the reliable automated recognition of facial expressions becomes a complicated task. The main intention of this paper is to develop an automated facial expression recognition system. To attain the goal, in this work, an improved particle swarm optimization bat tracking search algorithm optimized quantum convolutional neural network (IPSOBSA-QCNN) approach is proposed to accurately identify and classify facial expressions such as fear, happy, disgust, sad, surprise, negative, positive, and dominated. Although classification performance is limited by hyperparameter tuning concerns, the QCNN system efficiently recognizes face expressions at a high processing speed. Therefore, the hyperparameters of the quantum convolutional neural network (QCNN) are optimized to enhance the classification accuracy using the improved particle swarm optimization backtracking search algorithm (IPSOBSA). The proposed IPSOBSA-QCNN approach is evaluated using three diverse datasets namely the real-world affective faces dataset, the Emotic dataset, and the facial expression comparison dataset. The effectiveness of the proposed approach is determined by comparing the proposed IPSOBSA-QCNN approach with other existing methods using different evaluation metrics. The experimental results display that the proposed IPSOBSA-QCNN approach achieves higher classification accuracy of about 98% for the real-world affective faces dataset, 97.5% for the Emotic dataset, and 97% for the facial expression comparison dataset. [ABSTRACT FROM AUTHOR]

Position detection is essential for precise contactless manipulation as it can play an important role in analyzing the behavior patterns and motion regularities of levitated objects. However, traditional detection methods have several limitations, such as finite feature representation, low detection accuracy, and poor adaptability, notably for small targets. To address these issues, this paper proposes an effective detector called RFRM-CenterNet, which is the first attempt to detect the location of levitated objects based on a convolutional neural network. First, a receptive field refinement module consisting of multi-stage parallel dilated convolutional layers is designed to detect small levitated objects. Then, to increase the feature representation capability, a feature fusion network is designed, which employs a parallel fusion of ResNet50 and a receptive field refinement module. In addition, an experimental system is constructed, and a dataset is established to train the model and verify the performance of the proposed method. The experimental results show that RFRM-CenterNet has better performance in detecting levitated objects than other detection methods. [ABSTRACT FROM AUTHOR]

In free-form surface reconstruction technology, the ultimate goal is to obtain a computer model of a free-form surface. This study discusses the inpainting method of arbitrary surface data based on geometric convolutional neural networks. Reverse engineering is a process of product design technology reproduction, that is, reverse analysis and research of a target product, to deduce and obtain design elements such as the processing flow, organizational structure, functional characteristics and technical specifications of the product to produce functional similar but not identical products. This paper realizes the noise point elimination of point cloud data by studying reconstruction technology of free-form surfaces using a geometric convolutional neural network model. Point cloud data without gross errors are provided for later reconstruction. High-precision arbitrary surface data are realized for surface reconstruction. The repair of arbitrary surface data is completed based on BP and RBF neural networks, and a certain degree of data supplementing for the incomplete point cloud data measured is realized. Finally, free-form surface reconstruction is realized. In reverse engineering, a flexible surface is converted into a rigid surface, and the data points are collected using contact measuring equipment. The influencing factors of reverse engineering are work efficiency, technical ability and the influencing factors of innovative design. The data points obtained in this study are processed by denoising, streamlining and filtering. The sorted data points are used to obtain an optimized mathematical model after curve and surface fitting and smoothing. The research in this study proves that the data points obtained using the reverse engineering method can be very well applied in subsequent work. Therefore, this data collection method can be applied to other instances. In the experiment, the maximum error between the output of the RBF network and the test data is 0.0086 when surface 2 is repaired. When the hybrid learning algorithm is trained with different sample sets, the average width is 0.24, and the number of iterations is 1000. The experimental results of repairing various surface defect data show that the data repairing method has good versatility, fast data repairing speed and high precision, so it has high practical value. [ABSTRACT FROM AUTHOR]

We propose an image super-resolution method (SR) using a deeply-recursive convolutional network (DRCN). Single-image super-resolution reconstruction technology is to reconstruct fuzzy low-resolution images into clearer high-resolution images. It is a research hotspot in the field of computer vision and image processing. In recent years, the attention mechanism has been successfully applied in image super-resolution reconstruction. However, the existing methods use the channel attention mechanism and the spatial attention mechanism separately, or simply superimpose them, which cannot effectively unify the adjustment effects of both, and the performance is limited. This paper proposes a method that can merge channel attention and spatial attention into pixel attention, which achieves more precise adjustment of feature map information. The pixel attention convolutional neural network method built on this basis can improve the quality of image texture detail reconstruction. We have been tested on five widely used standard datasets, the experimental results show that the method is superior to most current representative reconstruction methods, especially in terms of high-definition picture texture restoration. [ABSTRACT FROM AUTHOR]

In this paper, we present a lightweight and effective change detection model, called TinyCD. This model has been designed to be faster and smaller than current state-of-the-art change detection models due to industrial needs. Despite being from 13 to 140 times smaller than the compared change detection models, and exposing at least a third of the computational complexity, our model outperforms the current state-of-the-art models by at least 1 % on both F1-score and IoU on the LEVIR-CD dataset, and more than 8 % on the WHU-CD dataset. To reach these results, TinyCD uses a Siamese U-Net architecture exploiting low-level features in a globally temporal and locally spatial way. In addition, it adopts a new strategy to mix features in the space-time domain both to merge the embeddings obtained from the Siamese backbones, and, coupled with an MLP block, it forms a novel space-semantic attention mechanism, the Mix and Attention Mask Block (MAMB). [ABSTRACT FROM AUTHOR]

Dummy trajectory is widely used to protect the privacy of mobile users' locations. However, two main challenges remain: (1) Map background information has not been modeled by machine learning methods in existing schemes, and (2) it is difficult to generate a good quality dummy trajectory that is similar to the real one. Focused on these two challenges, in this paper, we propose a dummy trajectory generation scheme with conditional generative adversary network (GAN), where the map features are extracted using convolutional neural network, which is regarded as a prior restriction of conditional GAN. Then, the movement pattern of the real trajectory is deduced by an auto-encoder and is involved in the dummy trajectory generation. Our model is trained and evaluated with two real-world datasets. Experimental results demonstrate that our scheme addresses these challenges well and defends against various attacks effectively. [ABSTRACT FROM AUTHOR]

Electroencephalogram (EEG) signals suffer substantially from motion artifacts when recorded in ambulatory settings utilizing wearable sensors. Because the diagnosis of many neurological diseases is heavily reliant on clean EEG data, it is critical to eliminate motion artifacts from motion-corrupted EEG signals using reliable and robust algorithms. Although a few deep learning-based models have been proposed for the removal of ocular, muscle, and cardiac artifacts from EEG data to the best of our knowledge, there is no attempt has been made in removing motion artifacts from motion-corrupted EEG signals: In this paper, a novel 1D convolutional neural network (CNN) called multi-layer multi-resolution spatially pooled (MLMRS) network for signal reconstruction is proposed for EEG motion artifact removal. The performance of the proposed model was compared with ten other 1D CNN models: FPN, LinkNet, UNet, UNet+, UNetPP, UNet3+, AttentionUNet, MultiResUNet, DenseInceptionUNet, and AttentionUNet++ in removing motion artifacts from motion-contaminated single-channel EEG signal. All the eleven deep CNN models are trained and tested using a single-channel benchmark EEG dataset containing 23 sets of motion-corrupted and reference ground truth EEG signals from PhysioNet. Leave-one-out cross-validation method was used in this work. The performance of the deep learning models is measured using three well-known performance matrices viz. mean absolute error (MAE)-based construction error, the difference in the signal-to-noise ratio (ΔSNR), and percentage reduction in motion artifacts (η). The proposed MLMRS-Net model has shown the best denoising performance, producing an average ΔSNR, η, and MAE values of 26.64 dB, 90.52%, and 0.056, respectively, for all 23 sets of EEG recordings. The results reported using the proposed model outperformed all the existing state-of-the-art techniques in terms of average η improvement. [ABSTRACT FROM AUTHOR]

Most knowledge graphs(KGs) are large and incomplete graph-structure database, which can be completed by predicting miss links according to the existing knowledge. The mainstream method is knowledge graph embedding (KGE) which is designed to learn low dimensional embedding of entities and relations. However, knowledge graph embedding still faces two major issues: (1) How to generate more expressive embeddings? (2) How to solve semantic polysemy of entities in different relations? In this paper, we propose a novel KG embedding model, RIECN (Relation-based Interactive Embedding Convolutional Network), which achieves high-quality performance and shows some advancements in modeling complex relations. In RIECN, FIR (Feature Interaction Reshaping) method is introduced to increase the feature interactions between entity and relation embeddings to generate more expressive feature maps. In addition, a new method of generating relation-based dynamic convolution filters, RDCF, is proposed. RDCF generates specific relation and hybird-size convolution filters, which enriches the feature maps of each entity improving the accuracy of link prediction task especially in complex relations scenario. We tested the performance of our model on five benchmark datasets. The experimental results show that the RIECN model significantly outperforms recent state-of-the-art models by 0.1–3.2% and 1.1–3.7%, in terms of MMR metric and Hit@1 metric, respectively. [ABSTRACT FROM AUTHOR]

Hyperspectral Image (HSI) is characterized by large number of bands with a high spectral resolution where continuous spectrum is measured for each pixel. This high volume therefore leads to challenges in processing the dataset. Objective of Dimensionality Reduction (DR) algorithms is to identify and eliminate statistical redundancies of hyperspectral data while keeping as much spectral information as possible. Combining spectral and spatial information offers a more comprehensive classification approach. Convolutional neural network (CNN) has the potential to extract complex spatial and spectral features embedded in Hyperspectral data. Wavelet transform belongs to the family of multi-scale transformation where the input signal is analyzed at different levels of granularity. Attention mechanism is a method in neural networks to guide the algorithm to focus on the important information in the data. In this paper, we use Multi-head Transformer-based Attention (Vaswani et al. in Attention is all you Need, http://arxiv.org/abs/1706.03762 2017) technique for Channel attention which captures the long-range spectral dependencies. The experimental results show that the proposed algorithm MT-CW Band Selection-based multi-head transformer for dimensionality reduction and Wavelet CNN-based algorithm for feature extraction yields impressive results in terms of information conservation and class separability. [ABSTRACT FROM AUTHOR]

Multi-stage hierarchical structure is a basic and effective design pattern in convolution neural networks (CNNs). Recently, Vision Transformers (ViTs) have achieved impressive performance as a new architecture for various vision tasks. However, many unknown properties of ViTs need to be further explored. In this paper, we empirically find that despite having no explicit multi-stage hierarchical design like CNNs, ViT models are able to automatically organize layers into stages (or blockgroups) to gradually extract different levels of feature information. Moreover, ViT models organize more highly similar Transformer blocks in the last stage, where the multi-head self-attention becomes less effective to learn useful concepts for feature learning and thus may limit the model to get the expected performance gain. To this end, we further recast a new ViT framework, named TransMCGC, replacing the inefficient Transformer blocks in the last stage of Vision Transformer with the proposed convolutional operation MCGC blocks. The MCGC block consists of two sub-modules in parallel: Multi-branch Convolution module to integrate local neighborhood features and multi-scale context information, and Global Context module to capture global dependencies with negligible parameters. In this way, the proposed MCGC block integrates collaboratively convolution locality and global dependencies to enhance the feature learning ability of the model. Finally, extensive experiments on six standard small-scale benchmark datasets, including CIFAR10, CIFAR100, Stanford Cars, Oxford102flowers, DTD and Food101, demonstrate the effectiveness of the proposed MCGC block and indicate that our TransMCGC models achieve better performance over baseline model ViT, while achieving competitive performance compared to state-of-the-art ViT variants. [ABSTRACT FROM AUTHOR]

Since wastewater treatment processes (WTP) are generally accompanied with intense coupling and fuzziness, conventional biochemical mechanisms-based methods cannot comprehensively express the WTP due to limited computational ability. In response to the challenge caused by fuzziness, this paper proposes a hybrid control and prediction system for modeling WTP with the fuse of Activated Sludge model, Convolutional neural network and Long short-term memory neural networks (AS-CL) with knowledge and data-driven characteristics. Moreover, the activated sludge model is employed to model the wastewater treatment process based on the perspective of knowledge. Besides, the hybrid neural network that combines convolutional neural network and long short-term memory model is adopted to model the WTP from the perspective of data. Then, a multi-layer perception model is set up to realize collaborative awareness of data and knowledge. Lastly, the proposed AS-CL has been evaluated by a real-world data-set collected from a real sewage treatment plant. The results show that compared with typical existing methods, the proposal improves modeling efficiency. With the collaborative modeling scheme, influence from fuzziness on WTP can be reduced to some extent. Compared with five benchmark methods of the two evaluation indexes, the results of AS-CL show that the average performance of this method exceeds 7% of the baseline. [ABSTRACT FROM AUTHOR]

The application of deep learning techniques in hyperspectral satellite images (HSI) classification has led to a significant increase in accuracy compared to machine learning techniques. The research objective is now shifted toward gaining classification accuracy with less number of spectral bands and training samples. With this research objective, the recent development of the hybrid convolution neural network (CNN) has resulted in the improvement of the accuracy of land use land cover classification of hyperspectral satellite images (HSI) and also provided a new dimension to explore. This paper proposes a variant of the 3D-2D CNN Hybrid model to extract representational features from a different class of land use land cover using different receptive fields in a multi-stack arrangement. The proposed model focused on reducing the number of bands required for classification and training samples but on the other hand improving the diverse representational feature extraction. The paper also explores through experimentation the possible replacement of monotonic activation function ReLu using non-monotonic functions like Swish and Mish in 3D-2D CNN Hybrid models. The experiments are carried out using publically available HSI datasets with a special focus on small training datasets to establish the efficiency and robustness of the proposed model. [ABSTRACT FROM AUTHOR]

The classification of environmental sounds is important for emerging applications such as automatic audio surveillance, audio forensics, and robot navigation. Existing techniques combined multiple features and stacked many CNN layers (very deep learning) to reach the desired accuracy. Instead of using many features and going deeper by stacking layers that are resource extensive, this paper proposes a novel technique that uses only a single feature, namely the Mel-Frequency Cepstral Coefficient (MFCC) and just three layers of CNN. We demonstrate that such a simple network can considerably outperform several conventional and deep learning-based algorithms. Through parameters fine-tuning of the data input, we reported a model that is significantly less complex in the architecture yet has recorded a similar accuracy of 95.59% compared to state-of-the-art deep models on UrbanSound8k dataset. [ABSTRACT FROM AUTHOR]

Vehicle drivers driving cars under the situation of drowsiness can cause serious traffic accidents. In this paper, a vehicle driver drowsiness detection method using wearable electroencephalographic (EEG) based on convolution neural network (CNN) is proposed. The presented method consists of three parts: data collection using wearable EEG, vehicle driver drowsiness detection and the early warning strategy. Firstly, a wearable brain computer interface (BCI) is used to monitor and collect the EEG signals in the simulation environment of drowsy driving and awake driving. Secondly, the neural networks with Inception module and modified AlexNet module are trained to classify the EEG signals. Finally, the early warning strategy module will function and it will sound an alarm if the vehicle driver is judged as drowsy. The method was tested on driving EEG data from simulated drowsy driving. The results show that using neural network with Inception module reached 95.59% classification accuracy based on one second time window samples and using modified AlexNet module reached 94.68%. The simulation and test results demonstrate the feasibility of the proposed drowsiness detection method for vehicle driving safety. [ABSTRACT FROM AUTHOR]

To create a safe and secure cyberspace, many researchers have created and proposed Intrusion Detection Systems (IDSs). Different types of IDSs are either designed to address only a single kind of intrusion or multiple variants. In this paper, we proposed a new context-aware feature extraction method as a pre-processing step for Convolutional Neural Network (CNN)-based multiclass intrusion detection. Feature selection is also used to reduce the feature space and classification time. For attack-type classification, we used CNN which is generally well-known for its better classification ability on image recognition tasks. We have exploited this ability to develop an IDS that can identify different types of intrusions ranging from 4 to 12 types. In our study, we evaluated our proposed Context-aware Feature Extraction-based CNN IDS on different datasets with multiple classes where it was able to successfully improve classification accuracy compared to the models with no preprocessing steps and other existing methods. We carried out the performance evaluation and comparison using four different datasets, NSL-KDD, CICIDS2017, ADFA-LD, and ADFA-WD. These datasets have either host-based or network-based features, which is another area that we investigated in this paper. [ABSTRACT FROM AUTHOR]

Advanced persistent threat (APT) attack is a malicious attack type which has intentional and clear targets. This attack technique has become a challenge for information security systems of organizations, governments, and businesses. The approaches of using machine learning or deep learning algorithms to analyze signs and abnormal behaviors of network traffic for detecting and preventing APT attacks have become popular in recent years. However, the APT attack detection approach that uses behavior analysis and evaluation techniques is facing many difficulties due to the lack of typical data of attack campaigns. To handle this situation, recent studies have selected and extracted the APT attack behaviors which based on datasets are built from experimental tools. Consequently, these properties are few and difficult to obtain in practical monitoring systems. Therefore, although the experimental results show good detection, it does not bring high efficiency in practice. For above reasons, in this paper, a new method based on network traffic analysis using a combined deep learning model to detect APT attacks will be proposed. Specifically, individual deep learning networks such as multilayer perceptron (MLP), convolutional neural network (CNN), and long short-term memory (LSTM) will also be sought, built and linked into combined deep learning networks to analyze and detect signs of APT attacks in network traffic. To detect APT attack signals, the combined deep learning models are performed in two main stages including (i) extracting IP features based on flow: In this phase, we will analyze network traffic into networking flows by IP address and then use the combined deep learning models to extract IP features by network flow; (ii) classifying APT attack IPs: Based on IP features extracted in a task (i), the APT attack IPs and normal IPs will be identified and classified. The proposal of a combined deep learning model to detect APT attacks based on network traffic is a new approach, and there is no research proposed and applied yet. In the experimental section, combined deep learning models proved their superior abilities to ensure accuracy on all measurements from 93 to 98%. This is a very good result for APT attack detection based on network traffic. [ABSTRACT FROM AUTHOR]

The goal of steganalysis is to unearth a concealed message hidden inside an innocent carrier by steganography. Steganography in the hands of unlawful people can pose a great threat to society. Secret message is embedded as a noise (residue) which silently disrupts the statistical nature of the carrier without visual transformation. For the purpose of unearthing covert communication, modeling this noise (residue) is a crucial factor. In this paper, a hybrid deep learning framework is proposed with convolutional neural network (CNN) to emphasize the need for residual based investigation for digital image steganalysis. In order to model the noise residuals, five types of handcrafted features are extracted, in which the first four types rely on acquiring noise residuals based on filtering by linear and nonlinear filters, and the fifth type relies on residual acquisition method based on Empirical Mode Decomposition (EMD). The extracted features are categorized using a robust CNN with a new parallel architecture that is capable of learning intricate details from input features to classify it as cover or stego. Mostly CNNs are trained with raw images, but in the proposed method, the 1-dimensional residual features are stacked into 2-dimensional grid and are then used to train the CNN. Three spatial content-adaptive and four spatial non-content-adaptive algorithms are used to evaluate the performance of the proposed architecture. The experimental results show the versatility and robustness of the proposed hybrid architecture towards these algorithms when compared to existing state-of-the-art steganalytic methods. A practical issue prevalent in deploying steganalysis in real-world is 'mismatch scenario,' and based on the experimentation done in this paper, the proposed architecture performs well under stego algorithm mismatch and payload mismatch scenarios. [ABSTRACT FROM AUTHOR]

Search and rescue are essential applications of disaster management in which people are evacuated from the disaster-prone area to a safer place. This overall process of search and rescue can be more efficient if an automated system can quickly locate the human or area where rescue is required. To provide a faster and accurate search of those places, this paper proposes a novel approach to search and rescue using automated drone surveillance. In this paper, a complex scene classification problem is solved using the proposed 3DCNN model. The proposed model uses spatial as well as temporal features of the video for the classification of the scene as help or non-help in the natural disaster. Due to the unavailability of such kind of dataset, it is impossible to train the model. Therefore, it is essential to develop a dataset for search and rescue. The proposed dataset is a first and unique dataset for scene classification using drone surveillance. The major contribution of this paper is (1) a novel 3DCNN powered model for scene classification in drone surveillance, (2) to develop the required dataset for the training of scene classification model, and (3) particular swarm optimization (PSO)-based hyper-parameter tuning for getting the best value of multiple parameters used for training the model. Our hybridization of parameter tuning with PSO helps for the convergence of parameter values of proposed 3DCNN model, and the proposed scene classification model (3DCNN+PSO) is applied to the dataset. The proposed model gives an impressive performance to help situation identification with 98% training and 99% validation accuracy. [ABSTRACT FROM AUTHOR]

This paper proposes an acceleration technique to minimise the unnecessary operations on a state-of-the-art machine learning model and thus to improve the processing speed while maintaining the accuracy. After the study of the main bottlenecks that negatively affect the performance of convolutional neural networks, this paper designs and implements a discarding technique for YOLOv3-based algorithms to increase the speed and maintain accuracy. After applying the discarding technique, YOLOv3 can achieve a 22% of improvement in terms of speed. Moreover, the results of this new discarding technique were tested on Tiny-YOLOv3 with three output layers on an autonomous vehicle for pedestrian detection and it achieved an improvement of 48.7% in speed. The dynamic discarding technique just needs one training process to create the model and thus execute the approach, which preserves accuracy. The improved detector based on the discarding technique is able to readily alert the operator of the autonomous vehicle to take the emergency brake of the vehicle in order to avoid collision and consequently save lives. [ABSTRACT FROM AUTHOR]

In recent times, convolutional neural networks became an irreplaceable tool in many different machine learning applications, especially in image classification. On the other hand, new research about robustness and susceptibility of these models to different adversarial attacks has emerged. With the rise in usage and widespread adoption of these models, it is very important to make them suitable for critical applications. In our previous work, we experimented with a new type of learning applicable to all convolutional neural networks: classification based on missing (low-impact) features. In the case of partial inputs/image occlusion, we have shown that our new method creates models that are more robust and perform better when compared to traditional models of the same architecture. In this paper, we explore an interesting characteristic of our newly developed models in that while we see a general increase in validation accuracy, we also lose some important knowledge. We propose one solution to overcome this problem and validate our assumptions against CIFAR-10 image classification dataset. [ABSTRACT FROM AUTHOR]

The preservation of plant specimens in herbaria has been carried out for centuries in efforts to study and confirm plant taxa. With the increasing collection of herbaria made available digitally, it is practical to use herbarium specimens for the automation of plant identification. They are also substantially more accessible and less expensive to obtain compared to field images. In fact, in remote and inaccessible habitats, field images of rare plant species are still immensely lacking. As a result, rare plant species identification is challenging due to the deficiency of training data. To address this problem, we investigate a cross-domain adaptation approach that allows knowledge transfer from a model learned from herbarium specimens to field images. We propose a model called Herbarium–Field Triplet Loss Network (HFTL network) to learn the mapping between herbarium and field domains. Specifically, the model is trained to maximize the embedding distance of different plant species and minimize the embedding distance of the same plant species given herbarium–field pairs. This paper presents the implementation and performance of the HFTL network to assess the herbarium–field similarity of plants. It corresponds to the cross-domain plant identification challenge in PlantCLEF 2020 and PlantCLEF 2021. Despite the lack of field images, our results show that the network can generalize and identify rare species. Our proposed HFTL network achieved a mean reciprocal rank score of 0.108 and 0.158 on the test set related to the species with few training field photographs in PlantCLEF 2020 and PlantCLEF 2021, respectively. [ABSTRACT FROM AUTHOR]

Travel time forecasting has become a core component of smart transportation systems, which assists both travelers and traffic organizers with route planning, travel schedule adjustments, ride-sharing, navigation applications, and efficient traffic management. However, timely and accurate travel time forecasting still remains a critical challenge owing to the complex nonlinear and dynamic fluctuations of spatial–temporal dependencies. Also, spatial sparseness is a big issue in traffic forecasting, since adopting the implicit interactions between the close traffic regions leads to superficial characterization of spatio-temporal dependences. In this paper, we propose a new deep learning-based framework (STTG-TTE) that addresses these drawbacks and improves the travel time estimation. First, we build a geo-hashing algorithm for the data sparsity issue that incorporates fluctuations of nearby and distant traffic situations in terms of spatio-temporal dependencies. Second, a new spatio-temporal correlation modeling method is proposed to fully leverage large-scale spatial and temporal traffic patterns using temporal convolutional networks integrated with a gated multi-modality mechanism. Then, for external factors' representation, a new dual-gated Res-Net multi-modality-based module is proposed. Finally, we fuse these representations of multi-components dynamically and utilize the transformer model, which is conducive to learning intersections among these multiple factors for obtaining accurate prediction results. Experiments on two large-scale real-world traffic datasets from two different urban regions (Chengdu taxi-datsets and NYC-Bike datasets) demonstrate that the proposed model is superior to state-of-the-art baseline models. [ABSTRACT FROM AUTHOR]

Capturing time and frequency relationships of time series signals offers an inherent barrier for automatic human activity recognition (HAR) from wearable sensor data. Extracting spatiotemporal context from the feature space of the sensor reading sequence is challenging for the current recurrent, convolutional, or hybrid activity recognition models. The overall classification accuracy also gets affected by large size feature maps that these models generate. To this end, in this work, we have put forth a hybrid architecture for wearable sensor data-based HAR. We initially use Continuous Wavelet Transform to encode the time series of sensor data as multi-channel images. Then, we utilize a Spatial Attention-aided Convolutional Neural Network (CNN) to extract higher-dimensional features. To find the most essential features for recognizing human activities, we develop a novel feature selection (FS) method. In order to identify the fitness of the features for the FS, we first employ three filter-based methods: Mutual Information (MI), Relief-F, and minimum redundancy maximum relevance (mRMR). The best set of features is then chosen by removing the lower-ranked features using a modified version of the Genetic Algorithm (GA). The K-Nearest Neighbors (KNN) classifier is then used to categorize human activities. We conduct comprehensive experiments on five well-known, publicly accessible HAR datasets, namely UCI-HAR, WISDM, MHEALTH, PAMAP2, and HHAR. Our model significantly outperforms the state-of-the-art models in terms of classification performance. We also observe an improvement in overall recognition accuracy with the use of GA-based FS technique with a lower number of features. The source code of the paper is publicly available here https://github.com/apusarkar2195/HAR_WaveletTransform_SpatialAttention_FeatureSelection. [ABSTRACT FROM AUTHOR]

In recent years, the world has dramatically moved toward using the internet of things (IoT), and the IoT has become a hot research field. Among various aspects of IoT, real-time cyber-threat protection is one of the most crucial elements due to the increasing number of cyber-attacks. However, current IoT devices often offer minimal security features and are vulnerable to cyber-attacks. Therefore, it is crucial to develop tools to detect such attacks in real time. This paper presents a new and intelligent network intrusion detection system named APAE that is based on an asymmetric parallel auto-encoder and is able to detect various attacks in IoT networks. The encoder part of APAE has a lightweight architecture that contains two encoders in parallel, each one having three successive layers of convolutional filters. The first encoder is for extracting local features using standard convolutional layers and a positional attention module. The second encoder also extracts the long-range information using dilated convolutional layers and a channel attention module. The decoder part of APAE is different from its encoder and has eight successive transposed convolution layers. The proposed APAE approach has a lightweight and suitable architecture for real-time attack detection and provides very good generalization performance even after training using very limited training records. The efficacy of the APAE has been evaluated using three popular public datasets named UNSW-NB15, CICIDS2017, and KDDCup99, and the results showed the superiority of the proposed model over the state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

The explosive growth of geospatial data is increasing requirements for automatic and efficient data learning abilities. Many deep learning methods have been widely applied for geospatial data understanding tasks, such as road networks and geospatial object detection. However, the demands for more accurate learning of high-level features require the use of deeper neural networks. To further improve the learning efficiency of deep neural networks, in this paper, we propose an improved convolutional neural network named EDense. First, we use its dense connectivity to integrate a CNN with an extreme learning machine. Then, we expand the kernels in the convolutional layers to increase the width of the network model. Furthermore, we propose one-feature EDense (OF-EDense), which is a simplified version of EDense, to fit conditions in which the number of parameters is strictly limited. Finally, the experimental results fully demonstrate the strong learning ability and high learning efficiency of EDense. [ABSTRACT FROM AUTHOR]

Accurate prediction of short-term origin-destination (OD) matrix is crucial for operations in metro systems. Recently, some deep learning-based models have been proposed for OD matrix forecasting in ride-hailing or high-way scenarios. However, the metro OD matrix forecasting receives less attention, and it has different prior knowledge and complex spatiotemporal contextual setting; for example, the sparse destination distribution and the incomplete OD matrices collection in recent time intervals due to unfinished trips before the predicted time interval. This paper designs a deep learning approach for metro OD matrix prediction by addressing the recent destination distribution availability, augmenting the flow presentation for each station, and digging out the global spatial dependency and multiple temporal scale correlations in the mobility patterns of metro passengers. Specifically, it first proposes to complete the recent OD matrices by combining some empirical knowledge including the historical mobility pattern and arrival time distribution. Then, it learns the complementary spatiotemporal contextual features by embedding methods to enrich the station representation. Finally, it captures global mobility trend of metro passengers at each origin station through aggregating the trend of all other origin stations by self-attention mechanism since the mobility synchronizes among stations from spatial perspective. Three temporal convolutional networks are leveraged to extract three temporal trends in passenger mobility data, i.e., recent trend, daily trend, and weekly trend. Smart card data from Shenzhen and Hangzhou metro systems are utilized to demonstrate the superiority of our model over other competitors. [ABSTRACT FROM AUTHOR]

Convolutional neural network, a strong deep learning technique, is used to detect diseases and perform image processing, recognition, and disease classification. The neural network is a breakthrough in technology that can process large sets of images in both 2D and 3D. In this paper, a novel Deep CNN model named NPNet-19 is proposed to determine maize crop infections. The model's accuracy and robustness are tested using an enhanced dataset of 15,960 images to classify six disease classes and one healthy class. The primary data source for testing the model has been acquired from the maize fields situated in the state of Telangana. The model was trained using images from the Plant Village and Kaggle repositories. On training datasets, the learning model showed an accuracy of 97.51%, and on testing datasets, it achieved an accuracy of 88.72%. The proposed model is empirically compared with pre-trained models such as DenseNet-121, Inception V2, ShallowNet-8, and CNN-SVM, and it showed a 10.57, 1.74, 2.15, and 1.1% improvement in classification accuracy, respectively. When compared to transfer learning models including Modified LeNet, DICNN, SoyNet, Adaptive CNN, 9-layer, and 13-layered architectures, the proposed model showed a 10.12, 14.52, 8.17, 3.88, 7.25, and 3.39% improvement in classification accuracy, respectively. Classification Prediction Analysis also has been used for statistical analysis of prediction data on testing and training images using error metrics such as RMSE, MSE, and MAE. It is observed that the performance of the proposed model is better in the detection of real-time diseases and their classification. [ABSTRACT FROM AUTHOR]

Deep learning techniques, in particular generative models, have taken on great importance in medical image analysis. This paper surveys fundamental deep learning concepts related to medical image generation. It provides concise overviews of studies which use some of the latest state-of-the-art models from last years applied to medical images of different injured body areas or organs that have a disease associated with (e.g., brain tumor and COVID-19 lungs pneumonia). The motivation for this study is to offer a comprehensive overview of artificial neural networks (NNs) and deep generative models in medical imaging, so more groups and authors that are not familiar with deep learning take into consideration its use in medicine works. We review the use of generative models, such as generative adversarial networks and variational autoencoders, as techniques to achieve semantic segmentation, data augmentation, and better classification algorithms, among other purposes. In addition, a collection of widely used public medical datasets containing magnetic resonance (MR) images, computed tomography (CT) scans, and common pictures is presented. Finally, we feature a summary of the current state of generative models in medical image including key features, current challenges, and future research paths. [ABSTRACT FROM AUTHOR]

Early diagnosis of brain tumor using magnetic resonance imaging (MRI) is vital for timely medication and effective treatment. But, most people living in remote areas do not have access to medical experts and diagnosis facilities. Nevertheless, recent advancement in the Internet of Thing and artificial intelligence is transforming the healthcare system and has led to the development of the Internet of Medical Things (IoMT). An automated brain tumor classification system integrated with the IoMT framework can aid in remotely diagnosing brain tumors. However, the existing methods for brain tumor classification in MRI based on traditional machine learning and deep learning are compute-intensive. Deployment of these methods in the real-world clinical setup poses a serious challenge. Therefore, there is a requirement for robust and compute-efficient techniques for brain tumor classification. To this end, this paper presents a novel lightweight attention-guided convolutional neural network (AG-CNN) for brain tumor classification in magnetic resonance (MR) images. The designed architecture uses channel-attention blocks to focus on relevant regions of the image for tumor classification. Besides, AG-CNN uses skip connections via global-average pooling to fuse features from different stages. This approach helps the network extract enhanced features essential to differentiate tumor and normal brain MR images. To access the efficacy of the designed neural network, we evaluated it on four benchmark brain tumor MRI datasets. The comparison results with the existing state-of-the-art methods revealed the robustness and computational efficiency of the proposed AG-CNN model. The designed brain tumor classification pipeline can be easily deployed on a resource-constrained embedded platform and used in real-world clinical settings to quickly classify brain tumors in MR images. [ABSTRACT FROM AUTHOR]

The difference between the convolutional neural network and the ordinary neural network is that the convolutional neural network contains a feature extractor composed of a convolutional layer and a subsampling layer. With the development of society and economy, the pace of urbanization is accelerating, and the number and types of urban buildings are also growing rapidly. Digital management has put forward higher requirements for 3D reconstruction of urban buildings. Aiming at explanation the question that sharpness form are proetrate to be blea or bewildered in CNN-supported structure birth from lofty-separation airy conception, an optimise construction birth algorithmic rule is converse to increase the construction brink of proud-resoluteness atmospheric semblance and the twist projection. Remote sensing image target recognition, as the main research content in the current remote sensing image application field, has important theoretical significance and extensive application value. In recent years, deep learning has become an emerging research direction in the field of machine learning, and convolutional neural network is a deep learning model that has been widely studied and applied. More specifically, the construction brink is better by realm vary recursive filter out, and the better appearance is fed into the U-Net nerval netting for making. Afterward, in custom to plentifully take advantage of the sumptuous detail shape of buildings on supercilious sake picture, we tempt to en plot impair from the manage copy and pigeonhole supported on the origin U-Net edifice to increase the school data. These beauty spot can remarkably fortify the procurement of edifice hie viterbilt characteristic in eager and invert intense lore. Finally, construction essence is instrument by mechanical advantage the quotation intense characteristic. The trial effect of edifice extract from the Panjin City have demonstrated that for the hoagie-optimum pattern data with division of shade areas, the everywhere assortment propriety of buildings recognized by U-Net is above 80%, and the zenith everywhere assortment truth of the amended regularity extension 83%. In this paper, through the research on the application of convolutional neural network in the field of image segmentation, the problems of low segmentation accuracy, long time and high cost in the task of aerial image building image segmentation are solved to a certain extent. The detection and segmentation method of buildings in aerial images based on CNN can automatically detect and segment buildings, and can segment a large number of buildings in aerial images in batches. In scenarios with high segmentation efficiency requirements. [ABSTRACT FROM AUTHOR]

Computer-aided medical image segmentation has been applied widely in diagnosis and treatment to obtain clinically useful information of shapes and volumes of target organs and tissues. In the past several years, convolutional neural network (CNN)-based methods (e.g., U-Net) have dominated this area, but still suffered from inadequate long-range information capturing. Hence, recent work presented computer vision Transformer variants for medical image segmentation tasks and obtained promising performances. Such Transformers modeled long-range dependency by computing pair-wise patch relations. However, they incurred prohibitive computational costs, especially on 3D medical images (e.g., CT and MRI). In this paper, we propose a new method called Dilated Transformer, which conducts self-attention alternately in local and global scopes for pair-wise patch relations capturing. Inspired by dilated convolution kernels, we conduct the global self-attention in a dilated manner, enlarging receptive fields without increasing the patches involved and thus reducing computational costs. Based on this design of Dilated Transformer, we construct a U-shaped encoder–decoder hierarchical architecture called D-Former for 3D medical image segmentation. Experiments on the Synapse and ACDC datasets show that our D-Former model, trained from scratch, outperforms various competitive CNN-based or Transformer-based segmentation models at a low computational cost without time-consuming per-training process. [ABSTRACT FROM AUTHOR]

A convolutional neural network can be constructed using numerical methods for solving dynamical systems, since the forward pass of the network can be regarded as a trajectory of a dynamical system. However, existing models based on numerical solvers cannot avoid the iterations of implicit methods, which makes the models inefficient at inference time. In this paper, we reinterpret the pre-activation Residual Networks (ResNets) and their variants from the dynamical systems view. We consider that the iterations of implicit Runge–Kutta methods are fused into the training of these models. Moreover, we propose a novel approach to constructing network models based on high-order Runge–Kutta methods in order to achieve higher efficiency. Our proposed models are referred to as the Runge–Kutta Convolutional Neural Networks (RKCNNs). The RKCNNs are evaluated on multiple benchmark datasets. The experimental results show that RKCNNs are vastly superior to other dynamical system network models: they achieve higher accuracy with much fewer resources. They also expand the family of network models based on numerical methods for dynamical systems. [ABSTRACT FROM AUTHOR]

During recent years, maintaining scientific and medical databases has turned out to be a major challenge in the medical sector. The medical data regarding the patients comprise diverse diagnostic medical reports and features for such disease need to be recorded carefully for providing high-quality treatments. Therefore, this paper proposes a hybrid deep convolutional neural network (HDCNN)-based COOT approach to predict the risk of diseases. Initially, an improved crossover-based levy flight optimization algorithm (ICLFDO) algorithm is employed in selecting and processing the unstructured textual data. Then later, the HDCNN-COOT approach predicts the diseases more accurately. In addition to this, the classifier predicts whether the patient is at risk of disease in the future or not. The efficiency of the proposed model is evaluated by using a dataset obtained from University Hospital of Ludwig Maximilian University of Munich, Germany. The data collected in the data center include 29,477,035 data items from 36,082 patients. The proposed model is capable of using structured data and uses textual data of the patients. Finally, the higher classification accuracy and improved performance of the classifiers can be viewed using the experimental results conducted on five different datasets. [ABSTRACT FROM AUTHOR]

Although convolutional neural networks (CNNs) are widely used in image classification tasks and have demonstrated promising classification accuracy results, designing a CNN architecture requires a manual adjustment of parameters through a series of experiments as well as sufficient knowledge both in the problem domain and CNN architecture design. Therefore, it is difficult for users without prior experience to design a CNN for specific purposes. In this paper, we propose a framework for the automatic construction of CNN architectures based on ResNet, DenseNet, and Inception blocks and the roulette wheel selection method with a dynamic learning rate. Compared with the state of the art, the proposed approach has a significant improvement in the domain of image classification. Experimental evaluation of our approach including a comparison with the previous works on three benchmark datasets demonstrates the effectiveness of the overall method. The proposed algorithm not only improves the previous algorithm but also keeps the advantages of automatic CNN construction without requiring manual interventions. The source code of the our framework can be found at https://github.com/btogzhan2000/ea-cnn-complab. [ABSTRACT FROM AUTHOR]

Convolutional neural networks method is a commonly used traffic sign recognition method based on deep learning over recent years. However, traffic signs contain objects of different sizes. Since small objects occupy a small input image area, the features that can be extracted are less, and the detection difficulty is greater than that of medium and large objects, and it is still challenging to achieve high-speed and high-accuracy detection of all objects of different sizes at the same time. In this paper, a model for detecting traffic signs is proposed, namely CDFF and CDFF-s. The model contains the following four modules: (1) in the backbone part of the model, we apply an improved activation function FMish to increase training stability, (2) after the backbone of the model, we apply the DFb-SPP module to perform context and semantic fusion, (3) in the neck part of the model, we use the DFb module for feature fusion, which also reduces the number of parameters, and (4) in the head part of the model, we propose a loss function SCIoU, which is optimized for small objects and the model is converged faster. The experimental results on the general traffic sign datasets TT100K and LISA show that the proposed two models can achieve accurate small object detection without losing the detection accuracy of medium and large objects. In addition, excellent results are also obtained on the remote sensing dataset RSOD with similar object size distribution. Meanwhile, the detection speed is faster than YOLOv4, which can meet the accuracy and real-time requirements of automatic driving systems and assisted driving systems. [ABSTRACT FROM AUTHOR]

Significant improvement has been noticed in salient object detection by multi-modal cross-complementary fusion between Depth and RGB features. The multi-modal feature extracting backbone of existing networks cannot extract complex RGB and color images effectively, which limits the performance of salient object detection in complex and challenging situations. In this paper, a composite backbone network with a mutual attention-based distinguished window is proposed to enhance the salient region and minimize the non-salient region. The distinguished window based on the channel-wise, spatial, mutual, and feature-level attention is inserted in each encoder stage to enhance the saliency features. Finally, a novel self-learning-based decoder, which is capable of utilizing multi-level features is designed to get the accurately dense prediction. The multi-level fusion is guided by deep global localized features. The performance of salient object detection could significantly be enhanced in this way. The extensive comparative and ablation experiments for the proposed framework have been conducted on the seven publicly available datasets for visual saliency. Experimental results have illustrated the effectiveness of the proposed framework and show better performance in comparison with the closely related state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Event Detection (ED) is an essential and challenging task in Information Extraction (IE). Recent advances in neural networks make it possible to build reliable models without complicated feature engineering. Neural network-based models commonly regard event detection as a char-wise or word-wise labeling task, which suffers from the problems of long-distance information miss-capturing, discontinuous labeling errors, etc., between characters/words and event triggers, especially in languages without natural word delimiters such as Chinese. In this paper, we propose a novel multi-layer Residual and Gated-Based Atrous Convolution Neural Network (RG-ACNN) framework, which attempts to alleviate above-mentioned problems. Specifically, the ACNN is introduced in our model to expand the receptive field to obtain multi-scale context information to capture dependencies between long-distance information. While gated and residual mechanisms are both imported to ACNN to improve our networks' capability of the information filtering and aggregation. Besides, RG-ACNN performs event detection in a char-wise paradigm, where a novel "head-tail dual-pointer" labeled strategy is used to overcome the incomplete continuous labeling problem. Experiments on the ACE2005-CN and several standard benchmark datasets show that RG-ACNN significantly outperforms state-of-the-art (SOTA) methods. [ABSTRACT FROM AUTHOR]

Misclassification is a critical problem in many machine learning applications. Since even the classifier models with high accuracy (e.g., > 95%) still introduce some misclassification error, it may not be possible to rely on the output of a classifier. In this paper, we introduce trustable learning, which prompts the learning model to yield only the true output, thus avoiding misclassifications. Whenever the model cannot decide the output accurately, the learning model should indicate that there could be a misclassification error if it is forced to classify, and hence, it should reject to make a decision or defer it to a human expert. Therefore, we develop a methodology for trustable learning and apply it to artificial neural networks and show that it is possible to develop a classifier with 0% misclassification error. We propose a novel neural network architecture named WisdomNet that could provide zero prediction error by introducing an additional neuron named as conjugate neuron that would indicate whether the network is able to classify the data correctly or not. The WisdomNet architecture can be applied to any previously built model, and we have evaluated WisdomNet with several network architectures such as multilayer perceptron, convolutional neural network, and deep network on different data sets. The results show that the WisdomNet is able to reduce the classification error rate to 0%, while labeling the data is difficult to classify as 'reject' at a low percentage of within around 10%. [ABSTRACT FROM AUTHOR]

Automatic handwriting recognition is an important component for many applications in various fields. It is a challenging problem that has received a lot of attention in the past three decades. Research has focused on the recognition of Latin languages' handwriting. Fewer studies have been done for the Arabic language. In this paper, we present a new dataset of Arabic letters written exclusively by children aged 7–12 which we call Hijja. Our dataset contains 47,434 characters written by 591 participants. In addition, we propose an automatic handwriting recognition model based on convolutional neural networks (CNN). We train our model on Hijja, as well as the Arabic Handwritten Character Dataset (AHCD) dataset. Results show that our model's performance is promising, achieving accuracies of 97% and 88% on the AHCD dataset and the Hijja dataset, respectively, outperforming other models in the literature. [ABSTRACT FROM AUTHOR]

Knee ligament tear injury is frequent in many volleyball, football, basketball, and cricket players. In the past, various deep learning-based ACL tear detection schemes using knee magnetic resonance imaging (MRI) have been presented. It has shown challenges in ACL tear detection performance due to class imbalance issue arising due to uneven training samples and poor feature representation. This paper presents a simple and effective ACL knee ligament tear detection method with a convolutional neural network (ATD-CNN) to lessen the intricacy of the network. Further, the self-attention mechanism is used to improve the feature representation of MRI image information, neglect the irrelevant information in deep features, and enhance the classification accuracy. To diminish the class imbalance issue, generative adversarial network (GAN) is used to construct the synthetic database. The performance of the ATD-CNN with self-attention is assessed on the MRNet database using precision, accuracy, F1-score, and recall. ATD-CNN provides an accuracy of 90.10% for the original and 93.93% and augmented datasets. However, the ATD-CNN without attention mechanism resulted in 89.60% and 92.30% accuracy for original and augmented dataset. The proposed ATD-CNN model indicates that it can be utilized to detect ACL tears automatically and outperforms the existing schemes for tear detection. [ABSTRACT FROM AUTHOR]

Document classification is becoming increasingly essential for the vast number of documents available in digital libraries, emails, the Internet, etc. Textual records frequently contain non-discriminative (noisy and irrelevant) terms that are also high-dimensional, resulting in higher computing costs and poorer learning performance in Text Classification (TC). Feature selection (FS), which tries to discover discriminate terms or features from the textual data, is one of the most effective tasks for this issue. This paper introduces a novel multi-stage term-weighting scheme-based FS model designed for the single-label TC system to obtain the optimal set of features. We have also developed a hybrid deep learning fine-tuning network based on Bidirectional Long Short-Term Memory (BiLSTM) and Convolutional Neural Network (CNN) for the classification stage. The FS approach is worked on two-stage criteria. The filter model is used in the first stage, and the multi-objective wrapper model, an upgraded version of the Whale Optimization Algorithm (WOA) with Particle Swarm Optimization (PSO), is used in the second stage. The objective function in the above wrapper model is based on a tri-objective principle. It uses the Pareto front technique to discover the optimal set of features. Here in the wrapper model, a novel selection strategy has been introduced to select the whale instead of the random whale. The proposed work is evaluated on four popular benchmark text corpora, of which two are binary class, and two are multi-class. The suggested FS technique is compared against classic Machine Learning (ML) and deep learning classifiers. The results of the experiments reveal that the recommended FS technique is more effective in obtaining better results than the other results. [ABSTRACT FROM AUTHOR]

One of the significant aspects of our digital world is that data are literally everywhere, and it is increasing. On the other hand, the number of cyberattacks aiming to seize this data and use it illegally is increasing at an exponential rate, and this is the challenge. Therefore, intrusion detection systems (IDS) have attracted considerable interest from researchers and industries. In this regard, machine learning (ML) techniques are playing a pivotal role as they put the responsibility of analyzing enormous amounts of data, finding patterns, classifying intrusions, and solving issues on computers instead of humans. This paper implements two separate classification layers of ML-based algorithms with the recently published NF-UQ-NIDS-v2 dataset, preprocessing two volumes of sample records (100 k and 10 million), utilizing MinMaxScaler, LabelEncoder, selecting superlative features by recursive feature elimination, normalizing the data, and optimizing hyper-parameters for classical algorithms and neural networks. With a small dataset volume, the results of the classical algorithms layer show high detection accuracy rates for support vector (98.26%), decision tree (98.78%), random forest (99.07%), K-nearest neighbors (98.16%), CatBoost (99.04%), and gradient boosting (98.80%). In addition, the layer of neural network algorithms has proven to be a very powerful technology when using deep learning, particularly due to its unique ability to effectively handle enormous amounts of data and detect hidden correlations and patterns; it showed high detection results, which were (98.87%) for long short-term memory and (98.56%) for convolutional neural networks. [ABSTRACT FROM AUTHOR]

Knowledge about the subtypes of a disease critically affects clinical decisions ranging from the choice of therapeutic options to patient management. If the understanding of a disease is partial and the subtypes of the disease are not yet known, a traditional supervised approach becomes untenable for disease subtype classification. In these contexts, unsupervised methods for subtype discovery are essential. There has been very little prior work on the discovery of phenotypic subtypes based on imaging. Moreover, within that limited body of work, the discovered phenotypes often lack interpretability. In this paper, we present a data-driven approach to discovering interpretable imaging phenotypes in 3D image volumes. In particular, the phenotypes are discovered in a latent space learned through a 3D autoencoder. To interpret the discovered phenotypes, we learn a convolutional neural network to classify the phenotype label and present VolPAM (Volumetric Phenotype-Activation-Map) to interpret the latent phenotypes in terms of their imaging footprints. We present results and visualizations on datasets of Computed Tomography images as an example 3D imaging modality. The proposed methods can become a useful aid to further the understanding of the condition in question through phenotype discovery and interpretability in terms of distinct aspects of the discovered phenotypes. [ABSTRACT FROM AUTHOR]

With the deepening of the research on corporate social responsibility (CSR), CSR has become increasingly important to enterprises. It can affect consumers' willingness to purchase enterprise products. Corporate social responsibility activities are used by many multinational corporations as a competitive business strategy to build a favorable corporate image in the eyes of consumers. Social media, made possible by developments in web and mobile technology, ushered in a new era of advertising on digital platforms. In this context, it has become an important topic to study the relationship between the authenticity of online CSR communication messages and consumers' purchase intentions. This work proposes an algorithm based on PSO-1DCNN joint optimization to analyze the impact of online CSR information authenticity on consumers' purchase intention in social media on the Internet platform. First, this work uses one-dimensional convolutional neural network (1DCNN) to model the relationship between the two. This model uses multichannel convolution feature, BN and Dropout strategy to promote performance for the model. Secondly, this work designs optimization measures from inertia weight and learning factor to build an improved particle swarm optimization algorithm (IPSO). Third, this work uses IPSO to optimize the initial network parameters of 1DCNN to build IPSO-1DCNN. The model has stronger convergence ability and convergence speed, in addition, it also has stronger global optimization ability. Fourthly, systematic experiments are carried out for the IPSO-1DCNN designed in this work, and the experimental data verify the superiority of this method. [ABSTRACT FROM AUTHOR]

This paper presents an efficient deep neural network model for diagnosing Parkinson's disease from gait. More specifically, we introduce a hybrid ConvNet-Transformer architecture to accurately diagnose the disease by detecting the severity stage. The proposed architecture exploits the strengths of both convolutional neural networks and Transformers in a single end-to-end model, where the former is able to extract relevant local features from Vertical Ground Reaction Force (VGRF) signal, while the latter allows to capture long-term spatio-temporal dependencies in data. In this manner, our hybrid architecture achieves an improved performance compared to using either models individually. Our experimental results show that our approach is effective for detecting the different stages of Parkinson's disease from gait data, with a final accuracy of 88%, outperforming other state-of-the-art AI methods on the Physionet gait dataset. Moreover, our method can be generalized and adapted for other classification problems to jointly address the feature relevance and spatio-temporal dependency problems in 1D signals. Our source code and pre-trained models are publicly available at https://github.com/SafwenNaimi. [ABSTRACT FROM AUTHOR]

Today, investigating gait recognition as a biometric technology has become necessary, especially after the COVID-19 pandemic broke out in the world. This paper proposes a deep structure procedure for precise human identification from the walking manner quickly, as each human has a unique way of walking. The proposed generative adversarial network (GAN) generates gait images from pedestrians of the CASIA and OU-ISIR gait datasets for normalizing the number of frames in each class in both datasets. Then, we extract the gait features from the normalized datasets by the pretrained convolutional neural network (CNN) models; AlexNet, Inception, VGG16, VGG19, ResNet, and Xception, and balance them using the synthetic minority oversampling technique (SMOTE) to enhance the recognition results. Several feature selectors are selecting the best features, including particle swarm optimization (PSO), grey wolf optimization (GWO), Chi-square, and genetic models. Finally, the proposed deep neural network recognizes the gait images precisely. Several performance assessment measures were generated to assess the model's quality, including accuracy, precision, sensitivity, specificity, false negative rate (FNR), intersection of union (IoU), and time. Experimental results showed that the inception model with genetic feature selector performed better in all used datasets than the current studies and was more robust against any change, achieving 99.3%, 99.1%, and 99.09% accuracy in identifying CASIA-B, CASIA-A, and OU-ISIR datasets, respectively in low time. [ABSTRACT FROM AUTHOR]

The convolutional neural network has been proven to be a robust recognition and diagnosis model for modelling diseases. In general, a convolutional neural network consists of feature learning which is considered a vital process in terms of feature extraction and classifier algorithms used for decision-making. In recent years, various pooling operations considered the spine of the feature learning layer have been developed in the literature; unfortunately, few of them can be spread to tackle the overfitting problem. In this paper, a novel pooling approach termed whole image average pooling is proposed to hold over the most significant feature information on the learning layer. The proposed method consists of two main stages: in the first stage, the average of all the values of the input image is calculated, in the second stage, the input image is divided into regions that represent the feature set of the input image, and then, the values of the regions are calculated. While obtaining the values of these features, the average of all values of the input image is taken into account. The proposed whole image average pooling approach addresses the overfitting problem by processing all pixel values in the image data and is used to get the important feature information from the medical datasets. To evaluate the performance of the proposed whole image average pooling, four clinically applicable tumour datasets as medical datasets have been trained and tested on seven convolutional neural network models AlexNet, VGG16, ResNet18, GoogleNet, ResNet50, R-CNN, and Faster R-CNN. The experimental outcomes demonstrate that the proposed whole image average pooling consistently outperformed the state-of-the-art pooling operation in terms of accuracy, sensitivity, specificity, positive predictive value, false-positive rate, and F1-measure evaluation criteria in three datasets. [ABSTRACT FROM AUTHOR]

Electroencephalogram (EEG) signals show the electrical activity of the brain, which are one of the inputs of the brain–computer interface (BCI). The BCI provides the communication path between the brain and the computer. One of the critical applications of BCI is Motor imagery (MI). MI is a mental process that a person practices or simulates a particular movement without physically acting. BCI allows the person to communicate with their environment independently of peripheral muscles and nerves, using EEG brain signals by assistive devices such as wheelchairs, robotic arms, and computers. In this paper, a space learning concept is proposed for EEG motor imagery signal classification. Our innovation in the proposed method is to increase and then, reduce the data dimensions, which has led to learning the efficient space for signals classification. It is based on two techniques: Multi-Kernel Learning (MKL) and dimension reduction. The composite kernel is made of a combination of four kernels by The Heuristic MKL Algorithm. This algorithm uses heuristic rules to estimate the weight of kernels with high accuracy and very little computational complexity. The weight associated with each base kernel and its parameters is calculated by the Equilibrium Optimizer. Dimensions of data are reduced to avoid the curse of dimensions. In this step, the number of dimensions of reduced space and the mapping matrix are learned to reduce the dimensions of data linearly. We selected ELM, KNN, and SVM classifiers for classification. The BCI Competition dataset was used for evaluation, which consists of five subsets aa, al, ay, aw, av, and two classes of the right hand and right foot. The proposed method with the ELM was improved the average classification accuracy and standard deviation by 3.9% and 2.28, respectively, and achieved 91.4% accuracy. The lower standard deviation than other methods shows that our method is more robust than all other methods to subject variety. The proposed method is compared with twelve state-of-the-art methods and has shown higher accuracy than other methods such as the deep convolutional neural networks. The results show the superiority of the proposed method over other methods in the Wilcoxon signed test. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) have been widely used in image classification tasks and have achieved remarkable results compared with traditional methods. Their main advantage is the ability to extract hidden features automatically using local connectivity and spatial locality. However, CNN cannot be applied to tabular data, mainly due to the unsuitability of the tabular data structure to the CNN input. In this paper, we propose a new generic method for the representation of multidimensional tabular data as color-encoded images that can be used both for data visualization and classification with CNN. Our approach, named FC-Viz (Feature Clustering-Visualization), is based on user-oriented data visualization ideas, such as pixel-oriented techniques, feature clustering, and feature interactions. The proposed approach includes a transformation of each instance of the tabular data into a 2D pixel-based representation, where pixels representing features with strong correlation and interaction are adjacent to each other. We applied FC-Viz to ten multidimensional tabular datasets with dozens to thousands of features and compared its classification and visualization performance with a state-of-the-art tabular data transformation method. The evaluation experiments show that our approach is as accurate as the state-of-the-art, but with much smaller images resulting in much more compact and faster CNN models. [ABSTRACT FROM AUTHOR]