Artificial Intelligence in China

Nowadays, the deep learning-based methods have been widely used in the scene-level-based image classification. However, the features automatically obtained from the last fully connected (FC) layer of single CNN without any process have little effect because of high dimensionality. In this paper, we propose a simple enhancing scene-level feature description method for remote sensing scene classification. Firstly, the principal component analysis (PCA) transformation is adopted in our research for reducing redundant dimensionality. Secondly, a new method is used to fuse features obtained by PCA transformation. Finally, the random forest classifier applying to classification makes a significant effect on compressing the training procedure. The results of experiments on the public dataset describe that feature fusion with PCA transformation performs great classification effect. Moreover, compared with the classifier softmax, the random forest classifier outperforms the softmax classifier in the training procedure.

Although the use of the convolutional neural network (CNN) improved the accuracy of object recognition, it still had a long-running time. In order to solve these problems, the training and testing datasets were split at four different proportions to reduce the impact of inherent error. Using model fine-tuning, the model converged in a small number of iterations, and the average recognition accuracy of BWN test can reach 96.8%. In the segmented dataset, the recognition accuracy of the former was 4.7 percentage points higher than the latter by comparing color dataset and grayscale dataset, which proved that a certain amount of color features will have a positive impact on the model. The segmented dataset was 0.9 percentage points higher than the color dataset; it shows that the model focused more on features of contour and texture by eliminating the background of images. The experiments showed that the binarized convolutional neural network can effectively improve recognition efficiency and accuracy compared with traditional methods.

Accurate hand detection is a challenging task because of large variations of hand images in real-world scenarios. We present a simple yet powerful multi-scale fully convolutional network structure that yields fast and accurate hand detection on challenging VIVA dataset. The proposed model directly detects and locates hands in driver’s cab of various size, shape, appearance, and illumination in full images without time-consuming region proposal step. The simple model with the well-designed loss functions promotes the proposed approach to achieve very good hand detection results.

It is of necessity to formulate an overall UAV regulation scheme that covers each UAV’s path and task implementation in the operation of UAV cluster. However, failure to fully realize consistency with pre-planned scheme in actual task implementation may occur considering changes of task, damage, addition or reduction of UAVs, fuel loss, unknown circumstance and other uncertainties, which thus entail a simultaneous online regulation scheme. By predicting UAV’s 4D track, posture, task and demand of resources in regulation and clarifying data set in UAV cluster operation, quick planning of UAV’s flight path and operation can be realized, thus reducing probability of scheme adjustment and improving operation efficiency.

Aiming at the problem that the integrity evaluation of satellite navigation and its augmentation system needs to analyze the characteristics of positioning error data, this paper proposes to use the time series analysis method to analyze the characteristics of the positioning error data of BDS single-point positioning and differential positioning from multiple angles of self-correlation, extremity and thick tail, which can provide the basis for deducing the integrity risk value by establishing an accurate positioning error distribution model. The result shows that the self-correlation of BDS differential positioning error is significantly lower than that of single-point positioning error. And the error of the two positioning methods in vertical and horizontal directions shows the characteristic of thick tail. Compared with the normal distribution, the single-point vertical positioning error has the most serious thick tail, reaching 420.3% and the differential horizontal positioning error has the lightest thick tail, only 22.9%. This indicates that the differential positioning method has better positioning performance and integrity.

Aging has become a serious problem facing the whole world. Falling is the leading cause of injury and death in the elderly. This paper proposes a fall detection algorithm based on machine learning XGBoost and full-field positioning. Using the data of gyroscope and acceleration sensor, we exploit the “full-field positioning” to increase the dimension of input data and propose a method “maximum satisfaction rate” to mark and train the threshold of data. The experimental results show that this design has obtained high accuracy on falling detection and perfect balance between sensitivity and specificity.

Aerial image labeling plays an important part in the mapping of maps with high precision. The knowledge about the range and intensive degree of aerial building segmentation is necessary for urban planning. Fully convolutional networks (FCNs) have recently shown state-of-the-art performance in image segmentation. In order to get better aerial images segmentation performance, we use a method of combing FCNs with multi-scale features and attention model in order to carry out segmentation automatically in aerial images. Attention model gives each scale feature added extra supervision to achieve better segmentation. Here, U-net and FCN-8s are used as original semantic segmentation model to train with multi-scale images and attention models. The datasets use different proportions of Inria Aerial Image Labeling Dataset, including two semantic classes: building and not building. The results show that the semantic segmentation model combined with multi-scale features and attention model has higher segmentation accuracy and better performance.

For interfering signals overlap with normal signals in both time and frequency domain, it is difficult to detect them. Therefore, this paper proposes a novel bidirectional recurrent neural network-based interference detection method. By utilizing the ability of recurrent neural network of extracting the nonlinear features of the time series context, the model can get a prediction of following signal samples and calculate the difference between prediction signal and original signal to do interference detection. The proposed method can achieve a better sensitivity and determine the exact location of the complete interfering signal. In the experiment part, we demonstrate the efficacy of this method in multiple typical scenarios of time–frequency overlapped wireless signals.

Collision risk analysis is an important part of airspace safety assessment. Based on Event model, the lateral position deviation probability of military aircraft is calculated. Combining with the probability model of lateral position deviation of civil aviation aircraft, the frequency of military aircraft collision box passing through separation sheet in high slope circling training is calculated. A collision risk assessment model between training flying aircraft in military training airspace and aircrafts flying in civil aviation route is constructed. Through the simulation calculation of the lateral collision risk of military and civil aviation, the size, layout, and use suggestions of military high slope circling training airspace are obtained, which can provide reference for the airspace safety assessment.

In recent years, deep learning has become increasingly popular in various fields. However, the performance of deep learning on imbalanced data has not been examined. The imbalanced data is a special problem in target detection and classification task, where the number of one class is less than the other classes. This paper focuses on evaluating the performance of the deep support vector machine (DSVM) algorithm in dealing with imbalanced human target detection datasets. Furthermore, we optimize the parameters of the DSVM algorithm to obtain better detection performance. It is compared with the stacked auto-encoder (SAE) and the support vector machine (SVM) algorithm. Finally, numerical experimental results show that the DSVM algorithm can effectively capture the minority class.

An emergency is a sudden and harmful event. It is of great significance to quickly identify the event and reduce the harm caused by the event. In this paper, the current advanced recurrent convolutional neural networks (RCNN) are utilized, but the traditional model cannot effectively identify the event, and the accuracy rate is not good enough. In order to solve this problem, the recurrent neural network and activation function part of the traditional model are improved, and through experimental comparison, the optimal model in the training model is selected. Finally, the accuracy of the model is 90%, the recall rate is 92.55%, and the F1 value, a metric that combines accuracy and recall, is 91.26%, which proves that the improved model has good effects.

MapReduce is a model of processing large-scaled data with parallel and distributed algorithms at a cluster, and it is composed of two stages: a map stage for filtering and sorting data and a reduce stage for the operation of summary. We develop a model with two connected queues: one upstream queue for the data flow to access the mappers and one downstream queue for the data flow to access the reducers. Also, we analyze the latency of processing a large scale of data using queueing models, in consideration of the coupling effects between these two queues for map and reduce, respectively. Our analysis results on various datasets and with various algorithms show that the MapReduce framework can almost linearly speed up with increasingly more processors, and adding mappers is usually more efficient than adding reducers to reduce the latency when processing a large-scaled dataset.

As a primary big data processing framework, Spark can support memory computing to improve the computation efficiency. However, Spark cannot handle the situation of a heterogeneous cluster, in which the nodes have different structures. Specifically, a primary problem in Spark is that the resource allocation strategy based on the number of homogeneous processor cores cannot adapt to the heterogeneous cluster environment. To solve the above-mentioned problem, we propose a zone-based resource allocation strategy based on heterogeneous Spark cluster (ZbRAS) and implement such a strategy to improve the efficiency of Spark. We compare the proposed strategy with the native resource allocation strategy of Spark, and the comparison results show that our proposed strategy can significantly enhance the execution speed of Spark jobs in a heterogeneous cluster.

This paper focuses on the prediction of combat outcomes in a local battle during a game of StarCraft. Through the analysis of the initial state of the two armies and considering the influence of the terrain in StarCraft on the combat effectiveness of both sides, the concept of the terrain correction factor is introduced to establish a mathematical model. Secondly, using SparCraft is to simulate battles and generate data sets. Finally, the maximum a posteriori probability estimation (MAP) is used to train the previously established data set to complete the parameter estimation of the mathematical model.

As a primary data processing and computing framework, Spark can support memory computing, interactive computing, and querying in a huge amount of data. Also, it can provide data mining, machine learning, stream computing, and the other services. However, the strategy of allocating resources among isomorphic processors cannot adapt to heterogeneous cluster environment due to its lack of load-based task scheduling. Therefore, we propose a dynamic load scheduling algorithm for heterogeneous Spark clusters by regularly collecting load information from each of the cluster node. Such an algorithm can dramatically reduce the allocation of load to the nodes which are already heavily loaded and in turn allocate more task to the idle nodes, thereby speeding up the process of job allocation in Spark. The experimental results show that the proposed algorithm can dramatically improve the computation efficiency by dynamically loading among the nodes in a heterogeneous cluster.

In this paper, priority-based mobile edge computing technology is used to decide the task migration problem on telemedicine devices. Tasks can be adaptively assigned to mobile edge servers or processed locally according to the current network situation and specific task processing environment, thus improving the efficiency and quality of telemedicine services. The main work of this paper is in three aspects. Firstly, priority is set for urgent tasks, and non-preemptive priority queues are set on the server side of MEC, so that the average stay time of each priority can be calculated according to the current situation of priority tasks queuing. Secondly, in the process of task transmission to the server, the channel resources are allocated adaptively by priority-based twice filtering strategy, and the final task migration decision is obtained by auction algorithm. Thirdly, compared with the existing mobile edge computing task migration model, the priority-based task migration model greatly guarantees the real-time and high quality of telemedicine.

Traditional rumor detection methods, such as feature engineering, are difficult and time-consuming. Moreover, the user page structure of Sina Weibo includes not only the content text, but also a large amount of comment information, among which the sentimental characteristics of comment are difficult to learn by neural network. In order to solve these problems, a rumor detection method based on comment sentiment and CNN-LSTM is proposed, and long short-term memory (LSTM) is connected to the pooling layer and full connection layer of convolutional neural network (CNN). Meanwhile, comment sentiment is added to rumor detection model as an important feature. The effectiveness of this method is verified by experiments.

The main purpose of object detection is to detect and locate specific targets from images. The traditional detection methods are usually complex and require prior knowledge of the detection target. In this paper, we will introduce how to use convolutional neural networks to perform object detection from image. This is one of the important areas of computer vision. In order to build up to object detection, we first learn about how we can get the object localization or landmark by a neural network. And then I will give the detail of sliding windows detection algorithm and introduce how to use the convolutional implementation of sliding windows to speed up the process. Then we will introduce the transfer learning and how to prepare your own learning data for training networks.

Researches show that the amount of mine gas emission is influenced by many factors, including the buried depth of coal seams, coal thickness, gas content, CH4 concentration, daily output, coal seam distance, permeability, volatile yield, air volume, etc. Its high-dimensional characteristics could easily lead to dimension disaster. In order to eliminate the collinearity of attributes and avoid the over-fitting of functions, Lasso algorithm is used to reduce the dimension of variables. After low-redundancy feature subset is obtained, the best performance model is selected by 10-fold cross-validation method. Finally, the gas emission is predicted and analyzed based on public data from coal mine. The results show that the prediction model based on Lasso has higher accuracy and better generalization performance than principal component analysis prediction model,and the accurate prediction of gas emission can be realized more effectively.

The water meter is a device for measuring the amount of water used by each household. Remote meter reading is one of the main ways to solve the waste of human resources caused by regular manual door-to-door access to mechanical water meter readings. The current use of image acquisition and then accurate reading of the water meter image is one of the ways of remote meter reading. In this paper, the convolutional neural network is used to predict the reading area, and then the non-maximum suppression algorithm (NMS) is used to remove highly overlapping results from prediction region results to obtain the position of the reading area. The experimental results show that with using the method proposed in this paper in the actual application scenario, the IoU of the images of 1000 test sets are all above 0.8 and then combined with the three-layer BP neural network for character recognition, the accuracy rate reaches 98.0%.

The problem of multi-pattern matching for RDF graph is an extended problem of subgraph isomorphism, where Resource Description Framework (RDF) is a graph-based data model for information sharing on Web. In real world, concurrent execution of multi-queries is more realistic than single query. However, the problem about re-computations of common subgraphs always limits the time efficiency of matching processing. To solve this problem, an algorithm of multi-pattern matching for RDF graph is proposed, which can response to multiple queries through one traversal of RDF graph. The experimental results show that our algorithm can avoid the re-computations of common subgraphs and improve up to 70$$\%$$ of time efficiency than basic line algorithm.

Radio frequency (RF) fingerprinting technology has been developed as a unique method for maintaining security based on physical layer characteristics. In this paper, we propose the RF fingerprinting by extracting the parameter characteristics such as information dimension, constellation feature, and phase noise spectrum in the transmitted information when applied to the universal software radio peripheral (USRP) software-defined radio (SDR) platform. To achieve a great performance improvement of classification, not only the traditional support vector machine (SVM) classifier, but also the machine-based integrated classifier bagged tree and the adaptive weighting algorithm weighted k-nearest neighbor (KNN) are both discussed. It is demonstrated that the proposed method achieves good classification performance under different signal-to-noise ratios (SNR).

In order to suppress the echo signal of target body and translation in Micro-Doppler signal processing, the influence is analyzed based on the feature of echo signal. The method of quadrature reception and translation compensation is proposed to suppress the interference of target body echo signal and translation. Theoretical analysis and simulation show that this method can effectively suppress the interference caused by the echo signal of target body and translation.

Approximation models are widely used in engineering reliability analysis due to the enormously expensive computation cost of limit state functions. In this paper, the weighted least-squared support vector regression (WLSSVR) method is used for model approximation. Sequential modeling is also considered to reduce the training sample number. A WLSSVR method with great gradient point (GGP)-based sequential sampling strategy is established and tested. The results show that the proposed method improves the global approximation accuracy.

In this paper, the adaptive partitioning problem of data stream under sliding window is discussed. Gaussian restricted Boltzmann machine (GRBM) model supporting decimal input is proposed, which can be trained through iteration for data reconstruction subsequently. At the same time, a data stream adaptive block algorithm based on Kullback–Leibler divergence (KL distance) is proposed to compare the probability distribution difference in the sliding window. Then, obtain the predicted value by the distribution of the previous data and determine whether the KL distance is within the confidence interval, so as to realize the adaptive adjustment of the sliding window, and the divided of data stream.

In this paper, we propose an intrusion detection method based on a convolutional neural network (CNN) for intrusion detection systems. In designing a deep intrusion detection model, mainstream deep learning techniques such as dropout, Adam, and Softmax classifier are used, respectively. Firstly, plain text data is dimensionally corrected and converted into grayscale images. Secondly, the CNN model obtains the feature map by learning features. Finally, the feature map is input to the Softmax classifier to obtain the detection results. The method is implemented on TensorFlow and tested on KDDCUP’99 data set. The results show that the model proposed can obtain high detection accuracy rapidly and satisfy the real-time detection requirements of complex network systems.

Satellite communication has become an indispensable means of communication. As satellite communication constellations become more and more complex, performance evaluation for satellite design, networking and applications grows more and more important. This paper summarizes current research on performance evaluation of satellite communication system and proposes application prospect of neural network in performance evaluation of satellite communication system. Identifying key parameters, adjusting evaluation model adaptively and comparing different satellite constellations’ performance are supposed to be three key application areas.

With the development of artificial intelligence technology, intelligent processing technology, which is represented by deep learning theory, has been widely used in radar community. The precondition of its application is the support of a large number of training dataset. However, for the technical direction of marine target detection, open dataset is insufficient to meet application requirements. To this end, framework of marine target detection dataset for intelligent radar application is constructed in this paper. It consists of three core modules, namely the detection equipment and data acquisition system, data management and processing, and integrated display. At present, sea clutter property cognition dataset, clutter suppression and target detection dataset, target characteristics dataset and detection performance evaluation dataset are preliminarily obtained. Dataset management and property analysis software are also developed, so as to serve applications more conveniently.

In this paper, we propose a novel data augmentation method named valuable absence augmentation (VAA) in order to alleviate the overfitting and evaluate the influence of the pedestrian valuable parts for the network performance. Specifically, we first train a base convolutional neural network model and obtain the attention map of the pedestrian. Then, we use the attention map to generate new samples. Finally, original samples and new samples are combined to fine-tune the base network model. We conduct experiments on a large-scale pedestrian retrieval database, i.e., Market-1501. Experimental results show that the pedestrian valuable part has a crucial influence for the network performance and that the proposed method achieves better performance than other state-of-the-art methods.

With the increasing number of hypertension patients, the monitoring of blood pressure information becomes an important task. In this study, an end-to-end approach is proposed to estimate blood pressure from the pulse wave signal. In this approach, a normalized single pulse wave is the input of a neural network, which consists of the convolutional layers and the recurrent layers, then outputs the corresponding blood pressure. The convolutional layers consist of one-dimensional convolutional layers and depth-separable convolutional layers. The gated recurrent unit (GRU) is used in the recurrent layer. Finally, a dense layer is used to output estimated values of blood pressure. In comparison with previous approaches, the proposed method does not require complicated feature extraction. It is only necessary to input a single pulse wave into the neural network and blood pressure can be estimated. The proposed approach is tested in the multi-parameter intelligent monitoring in intensive care (MIMIC) dataset, and the average absolute error is 3.95 mmHg for systolic blood pressure and 2.14 mmHg for diastolic blood pressure. This result fulfills the international standard of blood pressure measurement, which shows the proposed approach is simple and effective. In practice, the proposed method is designed to obtain blood pressure information from pulse waves.

With the advancement of information society, a large amount of data, which is in the form of stream, has been produced in many fields. As a result of its extensive application in the fields of sensor networks, banking and telecommunications, data stream mining is obtaining more attention. One of the most challenging steps to learn from data stream is to react to concept drift, as most of the existing data stream algorithms only deal with abrupt or gradual concept drifts. The existing work of detecting concept drift is mostly based on the changing of error rate of single window, making it difficult to be universally applied to different types of concept drifts. A method of detecting concept drift is proposed in this paper based on Kolmogorov–Smirnov test (K–S test).

The unmanned aerial vehicle (UAV) communications in the millimeter-wave (mmWave) band have found a wide range of concerns recently. In order to improve the communication capacity of UAV cellular networks, a low-complexity beam optimization method for the hybrid beamforming system with uniform planar arrays (UPAs) is proposed in this paper. First, the target cellular cell is quantified in spatial domain and the equivalent channel model of quantified region is established. Then, the data rate is formulated and an ideal precoding vector is achieved in terms of the expected beam gain. At last, a hybrid precoding method based on dynamic dictionary learning orthogonal matching pursuit (DDL-OMP) algorithm is introduced for producing an optimized beam. Simulation results demonstrate that our proposed method outperforms the traditional ones in achieving considerable capacity with faster convergence speed.

It is urgent to develop artificial intelligence technology and extend the brain of warship commanders with intelligent machines, thus assisting battlefield situation cognition and decision making. Since the 1990s, ensemble learning has become a new research focus in the field of machine learning. Ensemble learning can improve the generalization ability of the classification algorithms, and achieve the effect of “1 + 1 > 2”. Whether the performance bottleneck of single classifier can be broken by ensemble learning, so as to improve the ability of operational assistant decision of warship? This is a topic worth to discuss. In this paper, we focus on whether the torpedo can find the warship after the warship launched a torpedo decoy, and transform it into the two classification problem in machine learning. First, the classification data set is generated through offline simulation. Secondly, select decision trees as base classification method and Bagging as the typical of ensemble learning, and then the performance improvement of ensemble learning over single classifier is analyzed under ideal conditions. Finally, label noise, small training dataset size is taken into account, and the comparison experiments between ensemble learning and single classifier are performed further. The experimental results verify the feasibility of applying the ensemble learning to operation assistant decision.

Artificial Intelligence (AI) has made great progress in recent years, and it is unlikely to become less important in the future. Besides, it would also be an understatement that the game has greatly promoted the development of AI. Game AI has made a remarkable improvement in about fifteen years. In this paper, we present an academic perspective of AI for games. A number of basic AI methods usually used in games are summarized and discussed, such as ad hoc authoring, tree search, evolutionary computation, and machine learning. Through analysis, it can be concluded that the current game AI is not smart enough, which strongly calls for supports coming from new methods and techniques.

An intelligent exhaust system has been designed based on the characteristics of practical and efficient. The system takes STM32F103C8T6 single-chip microcomputer as the control core and consists of four parts: acquisition terminal nodes, routing nodes, master node, and monitoring management software. Use SHT20 sensors to collect environmental data. To realize the function of the system’s data transceiver and the exhaust fan control, data exchange between STM32 and the wireless transmission module LORA has been adopted. The upper computer not only contains PC monitoring management software, but also contains WeChat small program, which makes the fan control more simple and convenient. The system has been applied in a cross-linked workshop. The practical application shows that the system is stable, accurate, and easy to operate.

In order to study the crack detection of ancient buildings in aerial remote sensing images, this paper proposes to preprocess remote sensing images by combining block DCT transform with SPIHT algorithm terminating at the threshold to retain the target information to the maximum extent and reduce the number of processed images. Then, ResNet model is used to detect the cracks in ancient buildings.

With the development of speech recognition technology, there are wide applications of intelligent voice around the world. In the paper, we propose an intelligent voice control system which can realize the functions of voice activity detection, speech recognition, speaker authentication, instruction analysis, and automatic answer. A new voice activity detection (VAD) algorithm and backtracking algorithm are proposed to improve system performance. Then, we test our system in our own silumation corpus, we create the system evaluation indexes and the text result is satisfactory.

As a specific application of analytical methods on marine radar big data, this paper introduces deep learning theory into the field of sea clutter parameters estimation. A reasonable deep neural network model is built to estimate the parameters of amplitude distribution models so as to overcome the drawback of traditional methods based on statistical theory. In the proposed method, histogram method is used to preprocess the data, then deep neural network is trained with constructed dataset, and finally, parameter estimation results are obtained using test dataset. Validation results with simulation data and X-band radar-measured sea clutter data show that, compared with traditional estimation method, the deep neural network-based estimation method can improve parameter estimation accuracy significantly.

The number of ships has increased in recent years and the requirements for maritime security protection have become more stringent. It is important to improve the prediction accuracy and efficiency of maritime target in order to sustain maritime security. According to the real-time, efficiency and accuracy requirements of maritime target trajectory prediction, a prediction model based on RNN network is proposed to realize maritime target trajectory prediction based on AIS data. This paper uses AIS data between Longkou and Dalian to conduct experiments, and compares the prediction effects of two network models RNN and LSTM on the maritime target trajectory. It proves that the RNN network model has higher prediction accuracy and stronger learning ability. Based on the experimental results, this paper analyzes the characteristics of deep learning in long-term prediction and historical data dependence at the same time.

The traditional industrial production site ventilator adopts manual management, which is cumbersome and cannot be dealt with in the first time when abnormal conditions occur. In view of the above situation and demands, an industrial ventilator monitoring system based on Android is designed and implemented. The system takes ESP8266 module as the core, collects smoke value of industrial environment through smoke sensor, establishes server with Raspberry Pi, and develops an app with Android studio as development tool. Through the visual app interface, the on-duty personnel can not only view the status and usage situation of ventilators in different areas in real time, but also remotely control the opening and closing of multiple ventilators and set the timing task of ventilators. In addition, the system also achieves abnormal automatic processing. Once the abnormal smoke value is detected, the ventilator will be automatically opened, which has strong flexibility and high application value.

As a disruptive technology, swarm robotics is developing everyday, and attracts great attention. However, swarm robotics face some challenges which hinder swarm robots from broader application. Blockchain technology can provide a basic credible information environment of swarm robots, expand its application. This paper discusses how blockchain technology can provide benefits to swarm robotics, give a basic structure of swarm robots-oriented blockchain conceptual model. Finally, limitations and possible future problems that arise from the combination of these two technologies are described.

With dynamic topology, insufficient wireless bandwidth and variable path quality, new IP based applications across intra-flight ad hoc networks are always challenged by the fact that transmission services cannot always meet the requirements of the businesses. After analyzing the current status of transmission method of mobile ad hoc networks, this paper proposes a multi-mode transmission method based on business classification for intra-flight ad hoc networks. And the simulation results indicate that, compared with traditional transmission protocol, this method increases the throughput of intra-flight ad hoc networks greatly, while meeting the requirements of various businesses, such as the small capacity and low latency, the large capacity and low jitter, and the high reliability.

The current research of the statistical resolution limit (SRL), which is based on the hypothesis test, usually takes Taylor expansion and approximation to get a linear model of general likelihood rate test (GLRT) and achieves an analytical expression of the SRL. In the way, one dimension (range, angle) and two dimensions (range-Doppler, angle-Doppler) SRLs have been explored. In this paper we dwell on the three-dimension (3D) SRL in range-angle-Doppler domain and discuss the factors which make effects on 3D SRL. Our theoretical and simulation results both demonstrate that the 3D SRL is the weighting square sum of three respective SRLs, which will throw some insights into the systemic design to improve the resolution ability.

Aiming at the fast recognition of local image scenes with clear semantics in high-resolution remote sensing images, such as ports, airports and oil depots, a visual feature representation method for remote sensing images based on CSIFT features and a scene semantic recognition method based on PLSA is proposed. Experiments on typical remote sensing image scenes fully verify the effectiveness of proposed method.

In this paper, we analyze the development tendency of the intelligentization of command and control system and explore the application advantages of brain-inspired intelligence over traditional artificial intelligence methods in the intelligentization of command and control system. We also present a new approach through which general intelligence and system intelligence can be realized in command and control systems, and provide the evolution path of the invocation pattern of intelligent algorithms in command and control system.

Complex electronic system has complicated principal and excess amount of single equipment and malfunctions, which makes it difficult to run quick failure diagnosis and thus security was low. On the basis of directed graph model and correlation model, this paper explores the integrative technology of testability design and failure diagnosis for complex electronic information systems. Using vehicle electronic information system as an example establishes functional directed graph fault model and analyzes its testability, which provides reference for quick fault diagnosis.

This paper is aimed at improving name entity recognition (NER) accuracy. We replace the traditional bidirectional long short-term memory network (BiLSTM) with a changed BiLSTM, and then uses a CRF layer behind the changed BiLSTM layer to add the probabilistic relation of different Chinese characters.

Google’s AlphaGo shocked the world by easily defeating Korean Go player Lee Shi-shi, thus setting off a new upsurge of artificial intelligence research and application. Currently, artificial intelligence technology is developing at a speed beyond imagination, and it has become the core and key to the leap-forward development of every industry. However, application of artificial intelligence technology also encounters many problems. Key problems of military application of artificial intelligence are analyzed emphatically, and solutions to those problems are introduced, as a reference to further research.

With the development of the civil aviation industry, the pressure carried by the air traffic management (ATM) system, which is the core of civil aviation operations, is also gradually increasing, and therefore, it is necessary to use the power of various emerging information technologies to build a smart air traffic management system, thereby improving operational efficiency while ensuring safe operation of air traffic management. In recent years, the research of smart air traffic management has become a hot topic in academic circles, supported by industry guidance units such as the Civil Aviation Administration and the Air Traffic Management Bureau, and scholars at home and abroad have conducted relevant explorations on smart air traffic management. The paper comprehensively analyzes the research results of smart air traffic management in recent years, and for cloud computing, big data, artificial intelligence, Internet of Things (IoT), mobile Internet, five most important emerging information technologies, the research progress of combining with smart air traffic management is analyzed, meanwhile, proposed the shortcomings of the existing research and the suggested research directions for the next step, provide reference for further research on smart air traffic management.

The accuracy of 4D trajectory prediction is the foundation of trajectory-based operation and decision support tools of ATC. Meteorological modeling, especially the wind modeling, is the key element for improving the precision of trajectory prediction. Through analyzing the wind data in GRIB (GRIdded Binary) format, wind speed and direction at each levels can be obtained. Meanwhile, the 4D trajectory prediction model is constructed based on the wind effect to aircraft movements, especially ground speed and heading of aircraft. Taking arrival flights to Shanghai Pudong International Airport as an example, a simulation is conducted through the comparison of prediction results and the radar information. And the simulation results indicate the effectiveness of the proposed approach.

In order to effectively identify the rumor information in the Weibo platform, we propose a combined model based on deep learning, which includes convolutional neural network (CNN) that incorporates the attention mechanism and combines with the neural network of long short-term memory (LSTM) to implement a microblog rumor detection method for the characteristics of user-content relations. Firstly, the convolutional neural network incorporating the attention mechanism is used to extract the fine-grained features of the user-content relationship. Secondly, the LSTM network is used for coarse-grained feature extraction. Finally, the extracted feature vectors are classified by the Softmax classifier so as to achieve a good effect of rumor detection.

A hypercube labeling model HyperD was introduced to solve the problem of excessive bandwidth consumption and long query time between various nodes. By analyzing and adopting different strategies to improve the routing query algorithm of HyperD, a relatively more efficient routing query algorithm is proposed, which reduces the cost of communication and accelerates the data transmission between nodes. In order to verify the effectiveness of the algorithm, experiments were performed on randomly selected HyperD networks of various sizes. The improved routing algorithm was compared with the known routing algorithms, the results show that the improved algorithm is more efficient, and it can find the optimal solution in valid time.

Recent years, resting-state functional magnetic resonance imaging (rs-fMRI) and complex network analysis are guiding new directions in studies of brain disease. Here, this study aims to discriminate mild cognitive impairment (MCI) from cognitive normal (CN) by computer-aided diagnosis. For each subject, the functional brain network was constructed by thresholding with sparsity. Features were extracted through network measures and selected via the least absolute shrinkage and selection operator (LASSO). Features of two forms, computed separately under each threshold and integrated into area under the curve (AUC), were extracted. With features in the form of AUC, the support vector machines (SVM) classifier achieved an accuracy of 87%, which raised to 90% with threshold of 0.3. LASSO without stability selection was examined, achieving similar performance in classification to LASSO with stability selection while being less time consuming.

In this paper, we evaluate the performance of person re-identification (Re-ID) baseline under different implementation details including resize ratio of the pedestrian image, batch size and basic learning rate. To this end, we employ ResNet-50 as the classification model by modifying the original FC layer and apply a classifier to compute identity prediction values. We perform amounts of experiments to assess the effects of these implementation details on Market-1501 and experimental results show that these implementation details are very important for person Re-ID.

Decision-making processes of Air Traffic Management (ATM) need to organize information from different data sources. For example, reroute requires four-dimensional (4D) trajectory and weather information. Future ATM Decision-Support Systems (DSS) are expected to perform trustworthy complex decision-making automatically. This paper proposes an approach to DSS in ATM based on an ontology including concepts and instances of trajectories and meteorology. Temporal and spatial relationships are then set in this ontology. Reasoning rules are also build to represent knowledge in DSS. A case study shows a reroute scenario during a thunderstorm. In this scenario, information of the flight and the weather is combined to support decision-making by the proposed approach.

For high-altitude control area, in order to assist the ground control personnel to monitor short-term flight conflicts in real-time, and to solve the complex flight conflicts that may occur in multi-aircraft from a global perspective, this paper first establishes a spatial digital grid model based on flight safety intervals, and reasonably simplifies the motion model. Secondly, the short-term reachable domain of the aircraft is transformed into grid coordinates, and the numerically dimensioned method is used to obtain the conflicting reachable domain grid coordinates. On the basis of conflict detection, dynamic programming method is used to select the mutually exclusive grid coordinates which are obtained by traversing grid coordinates, and finally the conflict resolution decision of each aircraft is determined by the performance index. The simulation results show that the proposed method can effectively solve the complex flight conflict problem and provide the resolution decision that meets the safety interval and performance constraints before the TCAS (Traffic Collision Avoidance System), while the algorithm time can meet certain real-time requirements.

In this paper, a new algorithm for anomaly classification combined with shallow texture features is proposed for radioactive musculoskeletal images. The classification algorithm consists of three steps. The radioactive image is first preprocessed to enhance image quality. The local binary pattern (LBP) features of the image are then extracted and merged. Finally, the merged dataset is sent to the DenseNet169 convolutional neural network to determine whether it is abnormal. The method presented in this paper achieved an accuracy of 79.64% on the musculoskeletal radiographs (MURA) dataset, which is superior to the method that does not combine texture features. The experimental results show that the shallow texture features of the combined image can more fully describe the difference between the lesion area and the non-focal area in the image and the difference between different lesion properties.

In order to fuse information and analyze correlation more efficiently and flexibly, an air traffic management (ATM) knowledge graph-based method is proposed to reorganize the information flexibly and manage the fusion process dynamically. After that, a breadth-first and depth-first search-based correlation analysis method is designed to find deeper correlations and improve the searching efficiency.

In recent years, the number of patients with lung cancer has risen steadily, becoming the first malignant tumor in men and the second malignant tumor in women. Researchers at home and abroad have found that pulmonary nodule-assisted diagnosis can detect pulmonary nodules early and effectively reduce the risk of lung cancer. Therefore, deep learning has become a new hotspot in the diagnosis of pulmonary nodules. The research content of this paper is as follows: In this paper, we extract features of lung nodules with geometric features, gray value features, texture features and use support vector machine (SVM) and extreme learning machine (ELM) to train and classify the lung nodules. The convolutional neural network (CNN) deep learning method was used to extract the features of CT images of lung nodules, to establish a characteristic model of CT images of pulmonary nodules, and to classify the benign and malignant lung nodules. This paper presents a method for computer-aided diagnosis of pulmonary nodules based on improved CNN. This method uses the convolutional neural network (CNN) to extract the features of CT images of lung nodules and establishes the feature model of CT images of pulmonary nodules. The multi-scale convolution kernel depth learning is used to prove the advancement of improved algorithms.

This paper focuses on modifying generalized Hamilton system-based minimax theory for solving the interference suppression of multi-machine power systems with coupling performance. The system determines the interference suppression controller and reduces the conservatism of the traditional method by calculating the worst interference degree. The results of simulation show that this method and control strategy can make the system state converge to the initial equilibrium point rapidly under the influence of large disturbance to effectively improve the transient stability performance of power system.

In recent years, because deep learning technology can effectively learn features from data, it has become a powerful technical means in the field of image recognition. Research on image recognition can better promote the development of artificial intelligence and computer vision. This paper has conducted research and review of this field, introduced its important development and application, and made an attempt to promote further development in this field. Firstly, this paper introduces the structure of the network, and then introduces the common structural model of deep learning with CNN. The technical methods of reducing overfitting method, neural network visualization technology, inception structure, and transforming input images are discussed. Finally, the problems that still need to be solved in this field and the future of deep learning are introduced. It is pointed out that distributed computing, bit number reduction, migration learning, image style transformation, image generation, etc., are further research directions in the field of image recognition.

The new military revolution in the new era is spurring the development of combat means and combat forces to unmanned, intelligent, and clustered direction and gradually forming new methods of warfare and winning mechanisms. Unmanned combat platforms represented by unmanned aerial vehicle (UAV), unmanned combat vehicle, unmanned surface vessel, and unmanned underwater vehicle have gradually become new research areas in which various military powers are committing to. Among the various unmanned combat platforms, UAVs achieve the fastest development and have the most operational applications. This paper gives a brief overview of UAVs and their applications and focuses on the advantages and problems of UAV and UAV cluster.

Satellites are playing an increasingly important role in communication fields. Many commercial satellite communication systems are designed or constructed now. This paper reviews the recent developments in commercial satellite communications and outlines the performance characteristics of several typical satellite systems, such as Iridium-NEXT, LeoSat, OneWeb, StarLink, and so on. Prospects for the development of satellite communications are expected. The world of satellite communications is hotting-up, and a wide variety of design options will be explored. The future of high-speed “space Internet” is increasingly bright.

Compressed sensing (CS) has been extensively studied in image processing; however, the ill-posed inverse problem in the decoder is complicated, still leaving room for further improvement. Denoising-based approximate message passing (D-AMP) is a fast CS reconstruction algorithm that can transform the complex reconstruction problem into a classic denoising problem. But the existing denoisers are favorable on images with special features, so the D-AMP algorithm which has a fixed denoiser cannot universally achieve the best reconstruction quality for all types of images. In this paper, we propose a CS image reconstruction framework which can adaptively select the proper denoiser, and a feature-aware denoiser-selection method which extracts features by fast Fourier transform (FFT). Results show that the proposed feature-aware adaptive denoiser-selection-based CS image reconstruction method can obtain the best reconstruction quality for different types of images.

In recent development of communication systems, increasing amount of mobile terminals and multimedia content will require more and more resources to satisfy the need of users. However, there is rare study about the evaluation of large-scale complex network, from which the shortage of the network could be discovered. In this paper, we discuss the AI-based network evaluation method, we propose a novel discovery and ranking system using deep learning to collect and evaluate the network influence factor, and then, the key factors will be discovered to detect the advantages and shortages of network elements, deployments and scale. The AI-based detection and evaluation system is running along with the LTE-A system-level simulation platform, the accuracy and the effectiveness are evaluated, system shortages are successfully discovered within 100 trainings.

In this paper, based on annotating objects’ information and reasoning semantically with ontologies, the strategy of Semantic Web of Things (SWoT) for Internet of Things’ intrinsic contradiction is extended, and machine learning is introduced so that a novel hierarchical structure and the dynamic relationship between entities are proposed on SWoT. Specific contributions: (1) Establish of vertical and hierarchical framework of SWoT; (2) Definite on composite ontologies and (3) construct computational framework of Agent-based dynamic relationship model. And this paper proposes the essence that SWoT should be attributed to the interoperability between its various objects, which constructs inter-objects’ certain dynamic relationships.

If there is a large error in the initial position of the tracking method, the tracking performance will decrease. This paper proposes an accurate locating algorithm for initial tracking target with fusion of object-like information, saliency information, and priori information to provide accurate initial tracking target for further tracking method. Firstly, the probability distribution of object-like information is calculated by Edge Boxes which is used to detect the object-like regions. Then, the probability distribution of saliency information is calculated by spectral residual method. And the priori probability distribution is calculated according to the target initial position determined by the target detection. Finally, these three probabilities are adaptively fused to build multiple clues fused probability distribution. The precisely located initial tracking target area is the Edge Boxes object-like box with the highest average probability of fusion probability distribution. Experiments show that the accurate location method in this paper can achieve accurate relocation of the target in the case of selection box drift and scale changes and has good real-time performance.

In recent years, with the rapid development of artificial intelligence (AI), intelligent license plate (LP) detection which marks vehicle LP position from the images has become one of the core technologies of intelligent transportation system. The traditional license plate detection algorithm has low accuracy in some scenes, such as rotation, snow or fog, weak light, distortion, and blur. Therefore, this paper proposes a model based on convolutional neural network to realize vehicle LP detection. The model reduces the number of convolution layers in the case of ensuring accuracy, thereby reducing the number of parameters and saving time overhead. Through experiments, our model achieved an accuracy of 98.8%, and the training time and detection time were much less than the compare models.

Computer-aided diagnosis (CAD) systems have been introduced for therapeutic institutions to help radiologists to diagnose lung cancer at the early stage. This paper works on the parallelization of preprocessing steps in the CAD systems. We first propose a parallelization scheme based on static data grouping. The 3D lung CT images are grouped in the form of 2D image slices, and the grouped data is parallelized. Then, a pipeline-based parallelization scheme is proposed. Mutually independent steps are distributed to the respective execution units, and the preprocessing operations are performed in a parallel pipeline manner. Finally, a bus-based parallelization scheme is proposed. Based on the pipeline, a control unit is added to dynamically assign computing tasks to the various execution units. We test our schemes in an existing CAD system with data from Lung Nodule Analysis 2016 (LUNA16) dataset, and analyze the performance of data testing.

The waypoint behavior is a method of trajectory tracking control for unmanned surface vehicle (USV). We take the waypoint behavior as the research object, choose speed, steering angle, capture radius, slip radius, lead distance, lead damper as features, and establish a prediction model for trajectory tracking control of USV based on deep neural network. The model effectively predicts the navigation effect and provides assistance and reference for the maneuvering decision of the USV. Test results show that the predicted results and the reference samples have same tendency. The proposed model can improve the performance of trajectory tracking control for the USV.

The rapid development of autonomous driving has aroused widespread interest in academia and industry. Due to vehicular mobility, it is not feasible to adopt the existing coordinated learning approach based on static topology directly. To solve this issue, we propose a coordinated learning approach based on dynamic graph and reinforcement learning to enable distributed learning of cooperative lane changing. The dynamic graph model is constructed by evaluating the driving risk between each pair of vehicles. The lane change decision making of the vehicles is guided by the global optimal action based on the dynamic coordinated graph. Experiments verify that the proposed approach can achieve the accuracy and safety of lane change decision making, and with the increase in the number of vehicles, the approach has good scalability.

Sea target detection is widely used in military and civilian fields. Because of the space-time non-stationary characteristics of high resolution radar sea clutter, traditional target detection methods have many limitations and are limited by the use of scenarios. In recent years, with the progress of deep learning in image classification tasks, a series of target detection methods based on deep learning have emerged. By applying these methods to target detection in radar sea clutter, high accuracy and good generalization can be obtained. However, there are many new problems when these target detection methods, mostly based on computer vision, are introduced to target detection in radar sea clutter due to different data forms and detection standards. This paper mainly discusses the effects of different pretreatment modes of target detection in sea clutter using classification target detection framework on training speed and detection accuracy.

Because of multiple reflection and scattering, the mixed pixels in hyperspectral images are actually nonlinear spectral mixing. Traditional endmember extraction algorithm is based on linear spectral mixing model, so the extraction accuracy is not high. Aiming at the nonlinear structure of hyperspectral images, a novel endmember extraction method for hyperspectral images based on Euclidean distance and nonlinear dimensionality reduction is proposed. This method introduces Euclidean distance of image into the nonlinear dimensionality reduction algorithm of local tangent space permutation to remove redundant spatial information and spectral dimensional information in hyperspectral data and then extracts the endmembers from the reduced data by searching for the maximum volume of the simplex. Experiments on real hyperspectral data show that the proposed method has a good effect on hyperspectral image endmember extraction, and its performance is better than that of linear dimensionality reduction PCA and original LTSA algorithm.

Different from traditional text, short texts are characterized by high dimensionality, sparseness, and large text size. At the same time, some existing clustering ensemble algorithms treat each clusters equally, which will lead to the problem that the clustering results are not good enough. To solve this problem, this paper proposed a short text clustering ensemble algorithm based on convolution neural network (CNN). Firstly, the word2vec model is used to preserve the semantic relationship between words and obtain the multi-dimensional word vector representation; secondly, the feature is extracted from the original vector combining with the CNN; thirdly, clustering methods are used to cluster vectors; and then finally, Gini coefficient is used to measure the reliability of clustering, and the final clustering ensemble is carried out.

With the widespread application of the global positioning service (LBS) technology in the road network environment, the k-nearest neighbor query problem of moving objects in the road network has become a research hot spot for many scholars. This paper introduces the research status of k-nearest neighbor query of moving objects in the road network, analyzes and summarizes the research results of scholars at home and abroad, and finally analyzes the challenges of k-nearest neighbor queries for moving objects on the road network in the future.

As an unprecedented technological innovation, the Internet of Things has its inherent contradictions. The Semantic Web of Things is the solution to the internal contradiction of the Internet of Things. As an Intelligent Internet of Things, the Semantic Web of Things transforms grammar matching into semantic matching, which enhances the essence of the Internet of Things. This paper introduces the background and research significance of the Semantic Web of Things and introduces the branch of the current Intelligent Internet of Things. The related definition is given and the semantic relational structured network model of the Semantic Web of Things is proposed. This paper provides a theoretical basis and research basis for further research on Semantic Web of Things.

Facial expression recognition is an important application in computer vision. Generally, features which are used for facial expression recognition are mostly based on geometric and appearance features of image. This paper presents a novel method to identify facial expressions which exploring eye movements data labels as auxiliary labels to construct classifier, a Strengthened Deep Belief Network (SDBN) in a united cycle framework is constructed. This framework is formed as strong classifier by multi-weak classifiers voted. Experiments on Cohn-Kanade database showed that the proposed method achieved a better improvement in the task of facial expression recognition.

Aiming at the characteristics of color prediction and color matching in color appearance model, a vision local adaptation mechanism based on Sigmoid function of artificial neural network is introduced on the basis of iCAM color appearance model, and a real image reproduction algorithm based on Sigmoid-iCAM color appearance model is proposed. Compared with the real image reproduction algorithm based on CIECAM02 and iCAM color model, this algorithm can effectively improve the image contrast and local details while restoring the color of tone-distorted image.

The accurate classification of network game traffic is the technical basis for the campus network resources and the refined management of student learning behavior. This article first uses DNS reverse resolution to capture the domain name characteristics of the captured network game traffic, uses the domain name feature to mark the game traffic, and builds the network game feature data set. Then, the network game data set is used to train the decision tree CART to optimize the classification model parameters. Finally, the optimized classification model is tested with Moore standard data set and this data set, respectively. The results show that the accuracy of network game traffic classification based on DNS domain name resolution can reach 94%, and its classification performance is better than Moore data set.