Search

Many IoT devices are presently in use without sufficient security measures. The vulnerability of these devices to malware highlights the necessity for effective methods to identify malicious activity while protecting data privacy. Federated Learning is a collaborative learning technique that can create a fair model while keeping the local data private. To this end, this paper presents the SIM-FED, a novel model for detecting malware in IoT devices. This model utilizes both deep learning and federated learning algorithms. The SIM-FED offers a privacy-preserving solution for IoT malware detection, without sharing data, enhancing security in distributed environments. The model utilizes a lightweight one-dimensional CNN with optimized hyperparameters, reducing preprocessing, and computational overhead. Various federated aggregation strategies are evaluated, and the FedAvg strategy is selected to integrate the results of local models in the proposed model. Moreover, the model's resilience against white-box and black box cyber-attacks is assessed, revealing minimal performance degradation and highlighting its robustness, which is often overlooked in previous studies. The performance of the SIM-FED is assessed by conducting a series of experiments using the IoT-23 dataset. The results show that the SIM-FED outperforms other deep learning and federated learning models in terms of all evaluation metrics, achieving a remarkable 99.52 % accuracy. [ABSTRACT FROM AUTHOR]

Short-term energy consumption forecasting is the foundation of anomaly detection, scheduling and energy-saving control in manufacturing system. Existing methods do not reasonably distinguish between energy consumption nodes, but rather adopt overall prediction, resulting in poor prediction of target nodes. Consequently, this paper proposes a short-term energy consumption prediction method for the Attention Mechanism based on spatio-temporal deep learning (ST-Attention). Convolutional Neural Network (CNN) is utilized to extract feature extraction to acquire local high-dimensional features; Local long-term trend features are extracted using Gate Recurrent Unit (GRU) to reduce the inaccuracy of prediction due to the mixing of raw data;Finally, the Attention Mechanism is utilized to dynamically assign the weight of energy consumption nodes attribute to effectively deals with the difference, anisotropy and correlation between energy consumption nodes. The attention mechanism fuses extracted temporal and spatial features through weight assignment to enable short-term energy consumption prediction for manufacturing systems. The actual data collected from the aluminum profile factory is used to verify that the ST-Attention model providing better performance for short-term energy consumption prediction of the manufacturing system. [ABSTRACT FROM AUTHOR]

Representing three-dimensional (3D) objects by multiple views has become a common solution to the problem of 3D object retrieval. It has gained excellent achievements because of its remarkable adaptability and flexibility. In this paper, we develop a multi-view-based Siamese convolutional neural network for 3D object retrieval. It consists of two sub-networks which have the same architecture and also share the same set of weights. First, we generate a set of RGB images and binary images for each 3D object to capture local and global features. Second, the two sub-networks take corresponding images as input and avoid camera constraint by using average fusion layers. The final compact descriptors are learned by integrating features of sub-networks and then used for retrieval. Experimental results on two benchmarks, the PSB dataset and ETH dataset, testify that our proposed method receives superior retrieval performance compared to state-of-the-art methods. [ABSTRACT FROM AUTHOR]

• Interpretable DL Model: Our unique deep learning model interprets modulation mechanisms, efficiently extracting time and phase features with minimal resources. • Efficient Performance: Our approach achieves top performance on RML2016 and RML2018 datasets with fewer model parameters. • Fastest Embedded Speed: Outperforming 8 DL-based methods, our approach demonstrates the fastest speed on Jetson Nano and Raspberry Pi systems. Automatic Modulation Classification (AMC) constitutes a fundamental technology for enabling automatic demodulation in Cognitive Communication Systems (CCS). Due to the size, weight, and power (SWaP) constraints of embedded computers employed in CCS, there are limited computational and memory resources. While deep neural networks possess strong feature representation and high accuracy recognition capabilities, they usually come with a high number of network parameters and high computational complexity, thereby reducing the real-time processing ability of CCS. Therefore, neural network structures intended for CCS must be lightweight and computationally efficient. In this paper, we propose a high-performance and resource-friendly network model based on an analysis of the modulation mechanism of communication signals. The network extracts phase features and short-time features sequentially using directional convolutional filters. Long short-term memory (LSTM) units are then used to extract long-term features, and only one fully connected layer is used for classification. Experiments with a standard dataset consisting of 11 communication modulation types demonstrate that our proposed model achieves an accuracy greater than 84.5%, even when the signal-to-noise ratio (SNR) is 0 dB, and the model has only 29187 parameters. On a Jetson Nano embedded platform, the model achieves a processing speed of up to 375366 in-phase and quadrature samples/s. Overall, the results suggest that our proposed approach is both lightweight and highly efficient, making it more suitable for CCS applications. [Display omitted] [ABSTRACT FROM AUTHOR]

With the rapid evolution of the Internet of Things (IoT), network advancement has significantly influenced the increasing number of devices and advanced enhancements linked with it. Indeed the increasing number prevalence and sophistication of emerging cyber-attacks have highlighted the necessity for designing robust security application. In this paper, the Remora-based Deep Maxout Network model is Proposed. Here, the input data is acquired and forwarded to the pre-processing phase, wherein the missing value imputation approach is employed for creating a complete dataset. Later, the pre-processed data is then subjected to dimension transformation from the transformed data; the Convolutional Neural Network features are extracted, followed by feature selection based on Canberra distance. Here, detection is carried out using a Deep Maxout Network whose weights and training parameters are modified using the Remora Optimization Algorithm. However, the proposed model has delivered superior results with a high testing accuracy of 0.945 [ABSTRACT FROM AUTHOR]

Diabetic retinopathy (DR) is an illness occurring in the eye due to increase in blood glucose level. Among people in the age group of 70, 50% of deaths are attributed to diabetes. Early identification and appropriate treatment can reduce the loss of sight in many DR patients. Once the symptoms of DR are recognized, the severity of the disease should be evaluated for administering the right medication. This paper focuses on the classification of DR fundus images according to the severity of the disease using convolutional neural network with the application of suitable Pooling, Softmax and Rectified Linear Activation Unit (ReLU) layers to obtain a high level of accuracy. The performance of the proposed algorithm has been validated using Messidor database. In the case of healthy images, images of stage1, stage 2 and stage 3 of diabetic retinopathy, classification accuracies of 96.6% and 96.2%, 95.6% and 96.6% have been achieved. [ABSTRACT FROM AUTHOR]

Abstract In this paper, we present a deep learning-based recognition algorithm to identify pulsars by observing data containing millions of candidates including radio frequency interference and noise sources. The dataset is obtained from the High Time Resolution Universe survey created and updated by the Parkes telescope. We investigate several effective single and combined features via simple logistic regression. To deal with the imbalanced dataset, we oversimplify the original dataset at different sampling rates, which is also one of the learning parameters. After training the pre-processed dataset via a convolutional neural network, we provide a cross-validated evaluation of all candidates. Results show that the deep-learning based recognition algorithm can identify the pulsar and radio frequency interference signals with high accuracy. The precision and recall of radio frequency interference are both 100%, and those of pulsars are 91% and 94%, respectively. [ABSTRACT FROM AUTHOR]

Abstract Convolutional neural networks (CNNs) have been a focus area of machine learning in recent years, and they are widely used in vision and speech processing because of their superior performance. However, CNNs are usually resource-heavy to ensure higher accuracy, i.e., an accurate network with millions of parameters requires high performance computing devices. This prevents the use of CNNs in resource-limited hardware. In this paper, we propose a novel CNN simplification method to prune feature maps with relatively low discriminability magnitudes, which can produce a simplified CNN with reduced computational cost. Specifically, we define the critical points among the discriminability values of feature maps in each convolutional layer, and use these critical points to easily find the best pruning number of feature maps. Our experimental results show that in each convolutional layer of the VGG model, 15.6% to 59.7% of feature maps can be pruned without any loss of accuracy in classification tasks. [ABSTRACT FROM AUTHOR]

• This research paper developed BERT-BiCNN based deep learning model to capture software requirements from the formal requirements. • Since, traditional classification model suffers from poor generalization problem. Therefore, proposed model can be easily generalized to other dataset as there is no need of pre-processing. • We evaluate the proposed model on PROMISE benchmark dataset that contains software requirements. Automatic identification and classification of software requirements helps developer in next release planning of software engineering. • The reported results indicate that proposed model outperforms other deep learning based classifiers. In the last decade, requirements classification has emerged as hot research topic in Requirements Engineering (RE). Early identification of software requirements helps the development team in the design of software systems. Manual identification and classification of these requirements is time-consuming and labor-intensive task. To address this issue, machine and deep learning techniques have been studied for automatic classification of requirements. Furthermore, an efficient word embedding representation of the input data is a major concern in automatic approaches. This research work presents a novel requirements classification model that integrates Bidirectional Encoder Representations from Transformers (BERT) with Bidirectional Long-short Term Memory (BiLSTM) with Convolutional Neural Network (CNN) layer called BERT-BiCNN . In BERT-BiCNN, BERT is used as word embedding layer that extract the full combination of contextual semantics. BiLSTM obtains contextual information. The CNN mechanism has been employed to reduce the dimensionality of feature space by selecting the important features. The effectiveness of BERT-BiCNN is evaluated on the PROMISE dataset. On comparative analysis, it can be concluded that proposed approach outperforms six recent deep learning based architectures in binary and multi-class classification. [Display omitted] [ABSTRACT FROM AUTHOR]

Recently, convolutional neural network (CNN) has achieved impressive results in object classification tasks. Through various modifications and careful design of the inner structure of CNN, its performance has already become human-competitive, according to recent reports. To some extent, the testing accuracy depends largely on the decision boundaries produced by classifiers, which classify different objects into specific feature spaces. Hence, the relationships among samples in the feature space are critical. However, the softmax loss function that is used in most CNN models, does not directly contain the relationship information. In this paper, we propose a novel loss function, named inter-class constraint loss function, that maximizes the distance between different classes. Together with softmax loss, we can obtain larger inter-class distances and smaller intra-class distances in CNN, thus significantly improving the accuracy in classification. We achieve substantial improvements for the SVHN, CIFAR-10 and CIFAR-100 datasets using our proposed loss function. [ABSTRACT FROM AUTHOR]

Rapid developments in technology have led to increasing numbers of video surveillance devices in the home environment. The importance of video privacy protection has spurred the development of various video privacy protection methods. This paper proposes a method for evaluating the degree of privacy protection for multilayer compressed sensing video. The combination of CNN convolutional neural network and RNN convolutional network was used to extract video spatial-temporal feature mapping visual privacy protection scores, and the same model was used to map video practicality scores through classifiers. Compared with previous methods, the proposed approach achieves a better prediction effect and generalization performance for videos. Finally, an association model is established between visual privacy protection score and practicability score. This model quantifies the relationship between these aspects, providing suggestions for practical application and enabling the evaluation of other video privacy protection methods. [ABSTRACT FROM AUTHOR]

A multi-sensor network usually produces a large scale of data, some of which represent specific meaningful events. For event-driven multi-sensor networks, event classification is the basis of subsequent high-level decisions and controls. However, the accuracy improvement of classification is always a challenge. Recently the deep learning methods have achieved vast success in many conventional fields, and one of the most popular deep architectures is convolutional neural network (CNN) which sufficiently utilizes partial features of the input images. In this paper, we make some analogy between an image and sensor data, then propose a CNN-based method to improve the event classification accuracy for homogenous multi-sensor networks. An variant of AlexNet has been designed and established for classifying the event by acoustic signals. The results indicate that this CNN-based classifier outperforms than k Nearest Neighbor ( k NN) and Support Vector Machine (SVM) methods on our data set with a higher accuracy. [ABSTRACT FROM AUTHOR]

As an important part of urban intelligent transportation system, predicting users' demand for taxi using offline GPS data has attracted interests in recent years. In general, the distribution of traffic flow in different areas of the urban is different. Furthermore, the characteristics of traffic flow in different location areas present different. Therefore, it is challenging to achieve accurate prediction of users' demand for taxi in different spatiotemporal scenes. This paper presents a combined multi-scale residual calibration network (MS-ResCnet) using residual calibration network and multi-scale fusion mechanism to predict users' demand for taxi. Specifically, mapping rasterization and time series division methods are used to convert vehicular GPS data into spatiotemporal images of traffic flow within continuous sub grid areas. Then, the proximity, periodicity, and tendency, as significant characteristics of spatiotemporal images of traffic flow in each sub grid area are extracted. Thus, we establish a multi-dimensional spatiotemporal characteristic perception scheme of traffic flow in each grid area. Moreover, the datum characteristics and calibration characteristics of spatiotemporal images are extracted through the dual channel ResCnet network. Through the deep stagger training network, the full fusion of multi-scale spatiotemporal characteristics is realized. The performance of MS-ResCnet model is evaluated and verified using public datasets. The simulation results show that the traffic flow prediction performance of MS-ResCnet model is better than that of traditional STAR model. The root mean square error (RMSE) of the proposed method outperform the STAR model around 2.57%. [ABSTRACT FROM AUTHOR]

Deploying high-performance one-stage object detectors on resource-constrained applications is a challenging task. This paper analyzes factors affecting the computational complexity of one-stage detectors and proposes a lightweight and efficient detection framework named LEYOLO. Under this framework, a series of efficient feature extraction modules and a novel channel attention module are designed to compose a lightweight backbone network for detection task. To efficiently combine the features extracted from the backbone, a lightweight multiscale feature fusion structure with a weighted fusion method is proposed to avoid the overhead of dimensionality reduction and downsampling. Further, two detectors (i.e., LEYOLOs and LEYOLOm) are developed based on this framework. Experimental results show that LEYOLO achieves state-of-the-art trade-offs between performance and complexity, given only small computational budgets. [Display omitted] • An efficient and lightweight one-stage detector framework is designed. • Efficient feature extraction modules and a novel dilated channel attention block are designed. • Building CNN modules based on Ghost modules can reduce computational complexity. [ABSTRACT FROM AUTHOR]

• We convert the generated signals to two-dimensional images and then apply them to the convolutional neural network for classification. • The simultaneous faults are studied, which is most challenging to detect and isolate. • The proposed method outperforms better than other previous methods in terms of speed and accuracy. This paper's main purpose is to present a fault detection and isolation approach in the circuits employing a convolutional neural network and spectrograms. Monte Carlo analysis is performed for each of the existing faults, and several sample signals are generated. Then, spectrograms for one-dimensional output signals are calculated by the short-time discrete Fourier transform, and the convolutional neural network is trained using the spectrograms. The contribution of this paper is twofold. First, we suggest the power spectrogram to generate the features and apply them to the convolutional neural network. Second, usually, more than one fault occurs in circuit elements. So we study the simultaneous faults, which are the most challenging faults to be detected and isolated. Simulation results show that the proposed method has better accuracy than the existing methods from literature, and the computational time and the rate of false alarms have also reduced. [Display omitted] [ABSTRACT FROM AUTHOR]

Traffic identification has become a challenging task in recent years. Recently, deep learning methods have been extensively studied for network traffic classification recently. Unfortunately, these models require a large amount of training data. Another challenge with most traffic classification methods is that the features must be extracted by an expert. In these methods, finding the desired features that lead to a better classification is very tedious and time-consuming. In this regard, an excellent solution to address this challenge is to use of deep learning methods that automatically extract features. This paper is an attempt to address these issues by using a combination of the convolutional neural network (CNN), the ant-lion meta-heuristic algorithm (ALO), and the self-organizing map (SOM) to create a traffic classification model that can accurately identify traffic types. The proposed method was able to identify encrypted traffic and distinguish between VPN and non-VPN traffics. The model was evaluated using the "ISCX VPN-non-VPN" dataset, and achieved 98% accuracy. [Display omitted] • Preprocessing traffic data using entropy and entropy variance. • Automated feature extraction from the hidden layers of the one-dimensional convolutional neural network (1D-CNN). • Efficient feature selection with high classification accuracy using Ant-lion optimization. • Classification of new instances using a fuzzy-SOM based-clustering. [ABSTRACT FROM AUTHOR]

Parkinson's disease (PSD) and essential tremor (ET) are oscillatory and rhythmic movements in the human body with similar characteristics and becomes challenging to identify it accurately. Thus, the chances of misdiagnosis are high. Researchers employed machine learning (ML) algorithms to accurately classify ET and PSD patients. This requires manual feature extraction that, without knowing their importance in prediction purposes, can be mitigated with automated feature engineering using deep learning (DL). So, in this paper, we propose a convolutional neural network (CNN)-based classification model with seven hidden layers and different filter sizes for the accurate classification of PSD and healthy control (HC) subjects. A flatten layer converts three-dimensional data to one-dimensional Tensor flow. Finally, the dense layer outputs the classification of PSD and HC patients based on tremor intensity to identify the PSD patient's risk at an early stage. It outperforms the traditional models with 92.4% accuracy of tremor classification. [ABSTRACT FROM AUTHOR]

• Develop a CNN algorithm consists of two parts of feature extraction and classification in one model which makes it a unique and practical model in the area. • Develop a model to find the highest effective features for the market demand prediction • Training the CNN hyper parameters using a heuristic model based on PSO. • Considering the very complexity of the market demand in the new competitive environment using a novel fuzzy-based PSO • Use the fuzzy membership concept with aid of PSO to change its parameters dynamically rather than considering constant values. Abstract The concept of market demand forecasting is a very precious and significant problem in the industry without which the generator companies would not be able to give a suitable offer and may experience big losses. This research introduces a novel deep learning based on convolutional neural networks (CNN) to learn the market behavior in details. The model comprises a weight update mechanism in the backpropagation process of CNN to enhance its performance. Moreover, a fuzzy particle swarm optimization (FPSO) based algorithm is used to enhance the CNN performance and find the most accurate predictions. The proposed FPSO would utilize the fuzzy set theory and membership function to update the weighting coefficients. To validate this proposal, real market demand data are used as the benchmark and the results are discussed in the paper. The simulation results prove the suitable performance of the proposed CNN-FPSO model. [Display omitted] [ABSTRACT FROM AUTHOR]

In recent years, convolutional neural networks (CNNs) have outperformed conventional methods in end-to-end speaker identification (SI) systems. The CNN training time is considerably long due to the need for large amounts of training data and high costs of computation and memory consumption. This paper proposes a new CNN for text-independent SI inspired by the VGG-13 architecture with fewer parameters but an acceptable accuracy. In addition to the proposed CNN, the time complexity and memory cost of network training can be reduced through offline feature extraction by using a short segment of each audio sample and online data augmentation. According to the results on Voxceleb1, the proposed system is more accurate than the other state-of-the-art methods in SI. Therefore, the proposed CNN improved the accuracy and decreased the training time. [Display omitted] • Proposes a new CNN for text-independent speaker identification (SI) inspired by the VGG-13 architecture. • Offline Log-mel spectrogram feature extraction by using a short segment of each audio sample. • Online data augmentation. [ABSTRACT FROM AUTHOR]

Effective and timely fault detection and status monitoring of the industrial production process is essential to fully guarantee the operational safety. However, massive multi-source heterogeneous data analysis is facing many challenges. This paper proposes a process monitoring model combined with a parallel deep learning algorithm and Principal Component Analysis (PCA) method. Firstly, PCA is applied to realize fault diagnosis and extract fault characteristic variables, thus abnormal conditions can be revealed. In order to reduce the complexity of processing massive data, a parallel deep learning framework consisting of multi-models of convolutional neural network (CNN) and Long-Short Term Memory (LSTM) neural network is proposed to effectively predict the target variable status. Comprehensive experiments are taken under two real-system scenarios, and comparisons are made against four traditional neural network models to demonstrate its practicability and effectiveness. The generated results clearly show that the CNN-LSTM parallel model combined with PCA outperform other popular models due to its merged advantages of accurate time series prediction and effective fault feature extraction. [Display omitted] • A parallel deep learning approach with feature extraction and long time series prediction. • Multi-source unsupervised faulty variable extraction and detection can be realized. • Successfully solve the distribution power grid fault diagnosis in multiple locations. [ABSTRACT FROM AUTHOR]

In this paper, a new hybrid method of change detection and classification is proposed for precise detection and classification of series arc faults (SAFs) in photovoltaic systems. An artificial neural network (ANN) structure is applied for change detection at the first stage, which is then incorporated together with four different convolutional neural network (CNN) models with various dimensions as classifiers for the discrimination of SAFs at the second stage. The models used in the proposed method are 1D CNN, 2D CNN, 3D CNN, and 2D-based images. A comparison of the proposed approach and the state-of-the-art methods has been carried out in terms of accuracy and computational complexity. For a thorough evaluation of the proposed method's performance, studies have been conducted in both simulation and practice, considering various possible scenarios which may emerge. To such an aim, alongside the records from actual measurements in practice, nine models of SAF are also employed for simulation. The results show that the proposed method satisfies principle criteria such as reliability, fault classification error, overfitting, and vanishing solutions. [ABSTRACT FROM AUTHOR]

As an essential subproblem in the face recognition field, occluded face recognition has received considerable attention in recent years. However, satisfactory recognition accuracy for occluded faces has not been achieved. Therefore, in order to solve this problem, this paper proposes a convolutional neural network that has multidimensional serial feature extraction modules for occluded faces and uses the deep learning method to improve the recognition rate. In order to improve the expression ability of the network, the method extracts features from two dimensions — space and channel — and built a multidimensional serial feature extraction module. By "multiscale" and "dependence" processing features, the generalization ability was significantly improved. Through the experimental test, the recognition accuracy of the multidimensional feature network (MFNet) reached 90.35%. This accuracy is 17.89% (72.36%) better than that of the traditional algorithm. Meanwhile, compared with other convolutional neural networks, this network has different degrees of improvement, which are 0.68% (ArcFace), 2.14% (Visual Geometry Group Face), and 4.83% (DeepID3). [ABSTRACT FROM AUTHOR]

The non-contact monitoring of vital signs, especially the Heart Rate (HR) and Breathing Rate (BR), using facial video is becoming increasingly important. Although, researchers have made considerable progress in the past few years, there are still some limitations to the technology, such as the lack of challenging datasets, the time consuming nature of the estimation process, and non-portability of the system. In this paper, we proposed a new framework for estimating HRs and BRs by combining a Convolutional Neural Network (CNN) with the Phase-based Video Motion Processing (PVMP) algorithm. The experimental results show that our approach achieves better performance. Meanwhile, we introduce a new challenging dataset with fewer constraints, such as large movements, facial expressions and light interference. In addition, we developed a new Android application, which works in real time and offline, based on a CNN for HR and BR estimations. [Display omitted] • A non-contact system is proposed to monitor heart rate and breathing rate. • Phase-based video motion processing is used to attenuate noise and magnify signals. • Convolutional neural network is used to estimate heart rate and breathing rate. • Mean absolute error reduces by 8.3% compared with state-of-the-art methods. • A new less-constrained dataset for the heart rate and breathing rate is proposed. [ABSTRACT FROM AUTHOR]

This paper proposes an intelligent system, called IDGADS, for detecting algorithmically generated domains in the early stages based on easy and automatic computable features of real domain name system (DNS) traffic quickly without investing time in reverse engineering and/or log monitoring or dependency on external information like WHOIS/DNS response. IDGADS is a supervised deep learning model, trained over 17M domains from the reputed sources. It is implemented in Python and served as a service over the cloud for free testing of the public in the form of a web app. IDGADS is capable of detecting malicious domains up to 99% accuracy. Till 17-April-2020, 1160963 domains have been tested, and it has detected 817069 DGA-generated domains by the users of different countries. Since IDGADS is developed to check DNS queries only, thus it can be installed as the first line of defence in security stack for validating DNS queries before sending to DNS server. [Display omitted] • An intelligent deep learning system for detecting algorithmically generated domains. • Extracted intrinsic features from domain names without any human intervention. • Proposed system as first line of defence system for analyzing real domain names as a web service. [ABSTRACT FROM AUTHOR]

Security is currently one of the top concerns in our society. From governmental installations to private companies and medical institutions, they all have to address directly with security issues as: access to restricted information quarantine control, or criminal tracking. As an example, identifying patients is critical in hospitals or geriatrics in order to isolate infected people, which has proven to be a non- trivial issue with the COVID-19 pandemic that is currently affecting all countries, or to locate fled patients. Face recognition is then a non-intrusive alternative for performing these tasks. Although FaceNet from Google has proved to be almost perfect, in a multi-face scenario its performance decays rapidly. In order to mitigate this loss of performance, in this paper a cluster based on the Neural Computer Stick version 2 and OpenVINO by Intel is proposed. A detailed power and runtime study is shown for two programming models, namely: multithreading and multiprocessing. Furthermore, 3 different hosts have been considered. In the most efficient configuration, an average of 6 frames per second has been achieved using the Raspberry Pi 4 as host and with a power consumption of just 11.2W, increasing by a factor of 3.3X the energy efficiency with respect to a PC-based solution in a multi-face scenario. [ABSTRACT FROM AUTHOR]

• As the pornographic sources have been proliferated on the Internet, it is needed to provide a system to detect and filter them accurately. • A new framework has been introduced to make use of the benefits of deep convolutional neural network and ensemble learning. • We show experimentally that the proposed model outperforms the state of the art. • The proposed approach may be used in different applications such as intelligent filtering of unconventional images or medical images analysis. In this paper, we introduce an approach to discriminate unconventional images and their intelligent filtering. As the target data to this issue are huge and consequently, a handling approach might potentially be a very time consuming one, one of the major challenges to be solved by this introduced approach is its ability for dealing with large-scale datasets. A deep neural network might be a good option to resolve this challenge. It can provide a good accuracy while dealing with huge databases. In the proposed approach, the new architecture is introduced using a combination of AlexNet and LeNet architectures. It uses convolutional, polling and fully-connected layers. The results are tested on two large-scale datasets. These tests show that the introduced architecture is more accurate than the other recently developed methods in identifying unconventional images. The proposed approach may be used in different applications such as intelligent filtering of unconventional images or medical images analysis. Image, graphical abstract [ABSTRACT FROM AUTHOR]

The similarity of gray characteristics and shapes of the defects, stems, and calyxes in apple images is a difficult challenge for automatic identification and detection of apple defects in the machine vision field. This paper attempts towards finding an automatic and efficient way to identify apple defects. An automatic detection method is proposed for apple defects based on laser-induced light backscattering imaging and convolutional neural network (CNN) algorithm. Laser backscattering spectroscopic images of apples are obtained using semiconductor laser. We take preprocessing steps to get the finest image dataset for CNN. An AlexNet model with an 11-layer structure is established and trained to identify apple defects. We analyze how well the model does with the recognition effects of apple defects. The proposed CNN model for the detection of apple defects achieves a higher recognition rate of 92.5%, and the accuracy is better than conventional machine learning algorithms. The method based on laser backscattering imaging analysis and CNN theory provides an idea and theoretical basis for efficient, non-destructive, and online detection of fruits quality. [ABSTRACT FROM AUTHOR]

• A new deep learning method is proposed and applied to detect arm fractures in X-rays. • The proposed method achieves the state-of-the-art average precision (AP) in arm fracture detection. • Experienced radiologists are invited to annotate nearly 4000 arm fracture X-rays to evaluate our method. In this paper, a novel deep learning method is proposed and applied to fracture detection in arm bone X-rays. The main improvements include three aspects. First, a new backbone network is established based on feature pyramid architecture to gain more fractural information. Second, an image preprocessing procedure including opening operation and pixel value transformation is developed to enhance the contrast of original images. Third, the receptive field adjustment containing anchor scale reduction and tiny RoIs expansion is exploited to find more fractures. In the experiments, nearly 4000 arm fracture X-ray radiographs collected from MURA dataset are annotated by experienced radiologists. The experiment results show that the proposed deep learning method achieves the state-of-the-art AP in arm fracture detection and it has strong potential application in real clinical environments. [ABSTRACT FROM AUTHOR]