Artificial Intelligence for Communications and Networks

Cell counting plays an important role in biomedical research. There are always some phenomena such as indistinct intervals and target adhesion in cell images, which leads to poor segmentation effect and therefore inaccurate counting. In view of this situation, based on image binarization technology, this paper proposed a rapid cell count method combining mathematical morphology and connected domain labeling in which the cell images can be grayed, USM sharpened, binarized, morphologically processed, and connected domain labeled, and ultimately the number of cells could be calculated. The experimental results show that this method can effectively complete the segmentation of sparse cell images and intensive cell images, and the counting error is less than 5%.

In order to realize high-precision DOA tracking in space, researches on two-dimensional DOA estimation have been conducted in recent years. The existing algorithms often need large snapshots for estimation accuracy, going against the fast solution. Considering the low sensitivity of DOA estimation algorithm based on compressed sensing theory to the number of snapshots and the correct estimation with less sampling data, a modified two-dimensional multitask compressed sensing algorithm based on SVD decomposition is proposed in this paper. This algorithm makes up for the drawbacks of existing compressed sensing algorithms in dealing with multi snapshot problem and reduces the unnecessary calculation. Simulation results show that the proposed algorithm can solve the off-grid problem in compressed sensing, and has better estimation performance than other algorithms under the condition of low SNR and few snapshots.

With the development of mobile Internet technologies, wireless communication is facing huge challenges under the explosive growth of multimedia data, e.g. video conferences, online education. This makes it difficult to guarantee the communication quality where communication resources (bandwidth, channel, etc.) are limited. In this paper, we propose an image enhancement method to transform blurred images into images with high perceptual quality. The proposed method serves as a post-processing part for communication systems and is incorporated into the receiver. Specifically, we learn the prior of high quality images using a collected dataset. We train a neural network to accomplish this task and adopt a multi-scale perceptual loss as the objective, which is more consistent with the quality of experience (QoE). To validate the proposed method, we train our model on a large dataset with both blurred images and high quality images. Experimental results show that, using a pre-collected dataset with high quality images, the proposed approach can effectively restore the blurred images.

Indoor map construction techniques based on geomagnetic signals can achieve better effects for constructing indoor maps, because indoor geomagnetic signals are uniquely representative. An Indoor Map Construction Method based on Geomagnetic Signals and Smartphones (IMC-GSS) is proposed. The magnetic trajectory data are collected through smartphones and crowdsourcing technology, the Dynamic Time Warping (DTW) is utilized to cluster the obtained magnetic trajectory data, and the trajectory fusion technology based on the affinity propagation algorithm is applied to fuse the trajectories belonging to the same cluster in the magnetic trajectory domain to obtain a relatively accurate indoor path. The experimental results show that constructed indoor fingerprint map is reliable as well as effective.

Deep learning has been developed rapidly and widely used over the last decade. However, the concepts of adversarial example and adversarial attack are proposed, that is, adding some perturbations to the input of a deep learning model could easily change the prediction result. Deep learning-based NLP (natural language processing) classification algorithms also have this vulnerability. DeepWordBug algorithm is an advanced algorithm for generating adversarial examples, which can effectively deceive common NLP classification models. However, this algorithm needs to modify too many words to cheat NLP classification models, which limits its applications. In response to the shortcomings of DeepWordBug algorithm, this paper proposes an improving method to improve DeepWordBug. Drawing on the idea of Textfooler algorithm, the improved DeepWordBug adopts the method of dynamically adjusting the number of modified words, limits the maximum number of modified words. The new algorithm greatly reduces the number of words that need to be modified while ensuring the accuracy of NLP classification models as around 30%. It also ensures better practicality while maintaining transferability.

The randomness of the traditional measurement matrix in compressed sensing is too strong to be implemented on hardware, and when compressed sensing is used for image encryption, the measurement matrix transmitted as a key will consume time and storage space. Combined with the sensitivity of the chaotic system to the initial value, this paper uses Logistic-Chebyshev chaotic map to obtain random sequences with fewer parameters and construct measurement matrix. To test the measurement performance of the chaotic matrix, compare it with the Gaussian measurement matrix and the Bernoulli measurement matrix in the same compression encryption scheme. Pixel scrambling operation is carried out on the compressed image to complete the final encryption step, and the encrypted image is obtained. The reconstruction algorithm adopts the orthogonal matching tracking method to restore the image. The experimental simulation results show that the chaotic matrix has more advantages than the other two random matrices in image quality, and the encryption and decryption time is shorter.

As now available methods or systems based on image retrieval and visual researchs are implemented in an indoor environment, their retrieval accuracy and real-time positioning still have their own limitations. For this reason, this paper designs a visual indoor positioning system based on densely connected convolutional network image retrieval. Combine visual positioning with DenseNet-based image retrieval method. The problem of excessively deep network layers caused by the original convolutional neural network in pursuit of high retrieval accuracy is improved. Under the advantage of ensuring the high accuracy of image retrieval based on depth features, the problem of low real-time positioning caused by the long training time of the convolutional network model is improved. The simulation results show the feasibility of the positioning method in indoor environment, and the comparison experiment verifies the improvement of accuracy and speed as well as the reliability of the method.

At present, coin circulation automation technology has been widely used in many aspects, so it is necessary to install coin recognition and detection devices in related equipment to prevent coin confusion. However, many current coin detection methods have high requirements for hardware,which increases costs and makes it difficult to install and use equipment in narrow spaces. In this paper, we propose a coin recognition method with low hardware requirements and high accuracy. The design is to take a picture of the coin, detect the image and match the template to distinguish three different coins in the fourth edition of RMB. Compared with other detection methods using the eddy current method, it is lighter and easier to assemble, and can be easily embedded in narrow places.

To deal with the problem of de-noising and enhancement of radar signal time-frequency images, a method of secondary generating time-frequency images by generative adversarial network is proposed. Firstly, time-frequency analysis is used to generate the time-frequency image of the radar signal as the original data set 1. Then, after learning the data set 1 by using the generative adversarial network, a new data set 2 is generated, and the data set 2 has de-noising and enhancement effects relative to data set 1. Finally, the validity of the data set 2 generated by the time-frequency image singular value feature is checked. Experiments on the time-frequency images of five common radar signals are carried out. The results show that the method is effective in time-frequency image de-noising and increasing sample diversity.

One of the most important issues in communication is how to effectively detect signals. Being able to correctly detect the required signal is the basis for the partner to correctly implement the signal reception, and is also the basis for non-cooperative parties to implement information countermeasures and signal interference. The nonlinear signal detection method makes full use of the characteristics of the nonlinear system, and can detect the low SNR signal by converting the change of the signal into the change of the system state. Chaos theory is one of the nonlinear signal detection algorithms, and Duffing oscillator is the most typical. In this paper, the basic theory of Duffing oscillator is studied firstly, and the weak signal detection method based on Duffing oscillator is analyzed. In order to achieve signal frequency detection under narrowband noise conditions, narrowband noise is introduced into the Duffing oscillator to create a new Duffing oscillator model. Then analyzes the model by Melnikov equation, and the state form of the Duffing oscillator different from the traditional theory is obtained, that is, the probability period state of the oscillator under narrowband noise. A new method for weak signal detection using Duffing oscillator under narrow-band noise conditions is proposed. The state of the oscillator is judged by the period state time ratio (PSTR) method. Subsequently, using MATLAB to establish a weak signal detection platform based on PSTR method, the feasibility of detecting weak signals based on PSTR method under narrow-band noise conditions is verified.

The Industrial Internet of Things is considered to be an important cornerstone of future industrial development and has broad development prospects. It is being deployed to the society on a large scale. Sensor nodes in the Internet of Things will inevitably face many challenges. Due to the limitations of the existing sensor nodes, the traditional energy measurement methods for wireless sensor networks have been difficult to meet. Therefore, this article proposes a new routing protocol for low-power and low-power networks (RPL). The routing measurement method EEM meets the requirements of low-power lossy networks for link quality and energy consumption, and then uses the modified simulation software to test the EEM routing measurement, and performs performance verification on the packet loss rate and network load. The experimental results show that EEM retains the requirements of ETX routing metrics for link quality, realizes the awareness of node energy consumption, and optimizes network load.

To improve the link reliability and solve the problem of long feedback delay, a joint equalization and Raptor decoding (JERD) algorithm is proposed for underwater acoustic communication. Compared with the existing approaches, the Raptor code is adopted. The Raptor code is consisted of LDPC code generated by Mackey-1A and weakened LT code, and Raptor decoding adopts the global-iteration algorithm. The detector is iteratively adapted by switching soft information between the equalization and Raptor decoding at the Turbo processing stage. Simulation results validate the feasibility and show the advantages of the proposed algorithm against the existing approaches.

This paper first introduces the development history and future development trends of computer numerical control (CNC) machine tools. In order to greatly improve the work efficiency, this paper proposes a new type of CNC machine tool system, the central control device and the machine tool are separated, which can carry out one-to-many centralized management. In addition, in order to achieve the goal of mobile management, this paper also proposes an embedded CNC handheld terminal program, which can assist the work of CNC machine tools and save labor costs.

Simultaneous wireless information and power transfer (SWIPT) is a promising technology to realize simultaneous information and energy transfer by utilizing radio frequency signals. It extends the life of wireless networks and is conducive to the realization of green communications. In this paper, a subcarrier allocation-based SWIPT is studied to transfer information and energy in different subcarriers of an Orthogonal Frequency Division Multiplexing (OFDM) communication system. To improve the SWIPT performance, we maximize the energy efficiency of OFDM communication system while satisfying the constraints including minimum harvested energy, target rate and transmit power budget. To obtain the optimal solution, we investigate a dual-layer iterative optimization algorithm from Lagrange dual function to solve the energy efficiency optimization problem. The simulation results show that the energy efficiency of the proposed scheme can be effectively improved.

Deep space exploration is one of the three major aerospace activities of mankind in the new century, and deep space exploration is inseparable from the research on technologies of deep space communication. In the future, the goal of human space exploration will be extended to more and farther stars, analyzing the basic characteristics of the deep space communication link channel, and studying the specific communication problems of the nearer stars will be the necessary basis for the research of deep space communication. Focus on the characteristics of the communication channel between Earth-Moon and Earth-Mars. According to the influential parameters of different communication link, the impact of the communication link of Ka frequency band and below is simulated and analyzed to clarify each range of loss and the effect of each parameter, and the relevant channel characteristics of deep space communication link are obtained.

Two types of MIMO antenna arrays are proposed, towards millimeter wave technology. The antenna resonances are at 77 GHz. This paper mainly discusses the decoupling of antenna in MIMO system. The first array antenna is microstrip feed patch MIMO antenna which unit is an elliptical patch. A decoupling branch is added in the middle to improve the isolation to less than −15 dB. The second MIMO array antenna is with a resonance frequency at 77 GHz which meets the Chebyshev distribution. We use series resonance feed, and 1 × 16 line array is used as the MIMO antenna unit. We increase the line array space to achieve an isolation of less than −20 dB in the antenna frequency band.

Adaptive Modulation and Coding (AMC) and power control are adopted to balance spectral efficiency and coverage facing the problem of spectrum shortage in recent years. Considering the Orthogonal Frequency Division Multiplexing (OFDM) scheme with high spectral efficiency and the Single Carrier-Frequency Domain Equalization (SC-FDE) scheme with wide coverage, this paper combines the switching of carrier schemes with AMC and power control to maximize system throughput. The proposal of hybrid carrier communication system based on Weighted-type Fractional Fourier Transform makes the integration of OFDM and SC-FDE possible, which solves the problem of smooth transition between the two schemes. This paper analyzes the relationships between coverage and spectral efficiency in the flat fading channel and gives suggestions to the user equipment of different carrier schemes. Then we have proposed the calculation strategy of power control parameters and the switching strategy of the carrier schemes and modulation and coding schemes under power control in the frequency-selective fading channel.

Mobility refers to the ability to conduct “seamless” communication with network entities whose network location constantly changes. This paper examines the mobility support problem in IP and Named Data Networking (NDN), and identifies two dimensions in the mobility support solution space: the host dimension and data dimension. Existing host dimension solutions have exhausted the available design choices, and have not been able to achieve new breakthroughs in performance. Recognizing this limitation, this paper proposes a novel knowledge dimension. In the knowledge dimension, two knowledge-driven mobility support approaches, Topology-driven Intermediate Placement (TIP) and Trajectory-driven Reachability Update (TRU), are proposed. These approaches exploit knowledge such as network topology and movement trajectory to tweak the network and network services for better overall mobility support performance. A cross-architectural quantitative evaluation framework covering two communication scenarios and 5 quantifiable metrics is proposed to evaluate mobility support performance. Experiment results show that the knowledge-driven approaches significantly improve mobility support performance, demonstrating the potential of the knowledge-driven vision for providing better mobility support.

Due to the shortage of spectrum resources, Non-orthogonal multiple access (NOMA) has been considered as a forward-looking technology to enhance the performance for space-terrestrial satellite networks. In this paper, the power allocation algorithm is proposed on account of Stackelberg game to apply in the space-terrestrial satellite NOMA networks. The terrestrial base stations (BSs) and satellites are leaders and followers, respectively. The alternative direction method of multipliers (ADMM) algorithm is applied in BSs layer and satellites layer to acquire optimal power allocation scheme. The results indicate that the system energy efficiency has great promotion by the proposed algorithm.

In the Ad Hoc network of AODV protocol, local node failure will occur when a node dies due to high cost or the network energy consumption is too large and the problem is serious. In this paper, based on the routing cost design, an optimized design scheme of AODV protocol (E-AODV) is proposed through the establishment and maintenance method. Simulation results show that this method improves the performance of AODV network, reduces the death process of nodes and reduces the energy consumption.

In order to make the OLSR routing protocol more suitable for the self-organizing network of UAVs, this article optimizes the transmission method of HELLO messages and TC packets based on the change of the MPR selection method of the OLSR protocol. In order to adapt to the high dynamics and low density of UAV self-organizing network, the relative moving speed and link transmission quality are taken as the selection criteria of MPRs, so that the stability of MPRs can be improved. At the same time, in order to alleviate the routing overhead and energy problems caused by the increase of HELLO data packets, this article monitors the changes in the network and changes the sending frequency of HELLO messages and TC packets to reduce routing without affecting network updates. The problem of overhead and energy consumption. The network simulation is performed under the PPRZM motion model, and the protocol optimization effect is judged by the comparison of the packet delivery rate, the average end-to-end delay and the routing control overhead.

Unmanned Aerial Vehicles (UAVs) assisted communications are promising technology for meeting the demand of unprecedented demands for wireless services. In this paper, we propose a novel framework for delay minimization driven deployment of multiple UAVs. The problem of joint non-convex three dimensional (3D) deployment for minimizing average delay is formulated and solved by Deep Q network (DQN), which is a reinforcement learning based algorithm. Firstly, we obtain the cell partition by K-means algorithm. Then, we find the optimal 3D position for each UAV in each cluster to provide low delay service. Finally, when users are roaming, the UAVs are still able to track the real-time users. Numerical results show that the proposed DQN-based delay algorithm shows a fast convergence after a small number of iterations. Additionally, the proposed deployment algorithm outperforms several benchmarks in terms of average delay.

Unmanned aerial vehicles (UAV) has become a popular auxiliary method in the communication field due to its mobility and mobility. The air base station (BS) is one of the important roles of UAV. It can serve the ground terminals (GTs) without being restricted by time and space. When GTs are scattered, trajectory optimization becomes an indispensable part of the UAV communication. In this paper, we consider a UAV-assisted network with underlaid D2D users (DUs), where the UAV aims to achieve full coverage of DUs. Trajectory planning is transformed into the deployment and connection of UAV stop points (SPs), and a K-means-based trajectory planning algorithm is proposed. By clustering DUs, the initial SPs is determined. Then add new SPs according to the coverage, and construct the trajectory. The simulation analyzes the validity of the algorithm from the distribution of DUs and the number of initial cluster centers. The results show that the proposed algorithm is compared favorably against well-known benchmark scheme in terms of the length of the trajectory.

In recent years, unmanned aerial vehicle (UAV) technology has been widely used in industry, agriculture, military and other fields, and its positioning problem has been a research hotspot in this field. To solve the problem of invalidation of integrated navigation of global positioning system (GPS) and strapdown inertial navigation system (SINS) in indoor and other areas, this paper presents a multi-source information fusion location algorithm based on machine vision positioning and SINS. Based on image coordinate system (ICS), body coordinate system (BCS) and navigation coordinate system (NCS), combined with AprilTags recognition and positioning technology, this paper builds NCS with AprilTags array to get the position observation of UAV. Based on the idea of multi-source information fusion, this paper applied third-order fused complementary filter algorithm, which combines with the SINS to obtain accurate three-axis speed and position estimation. Finally, the reliability is verified by the test of the UAV experimental platform.

In recent years, the capabilities of UAV systems have continued to improve, and they have emerged in military and civilian fields such as urban counter-terrorism reconnaissance, disaster monitoring, logistics distribution, and traffic diversion, and their application prospects are particularly broad. UAV positioning is a necessary link for UAVs to perform tasks and an important manifestation of UAV’s autonomous capabilities. How to meet the positioning requirements of UAVs in environments with weak or no GPS signals such as urban buildings/forests/indoors has become a research hotspot in the UAV field. This paper introduces several UAV positioning methods that can work in GPS denial environment, analyzes their advantages and disadvantages, the current challenges in UAV positioning and finally looks forward to the future development.

Industrial Internet of things (IIoT) has been envisioned as a key technology for Industry 4.0. However, the battery capcity and processing ability of IIoT devices are limited which imposes great challenges when handling tasks with high quality of service (QoS) requirements. Toward this end, in this paper we first use multiple unmanned aerial vehicles (UAVs) equipped with computation resources to offer computation offloading opportunities for IIoT devices due to their high flexibility. Then we formulate the multi-UAV-enabled computation offloading problem as a mixed integer non-linear programming (MINLP) problem and prove its NP-hardness. Furthermore, to obtain the energy-efficient solutions for IIoT devices, we propose an intelligent algorithm called multi-agent deep Q-learning with stochastic prioritized replay (MDSPR). Simulation results show that the proposed MDSPR converges fast and outperforms the normal deep Q-learning (DQN) method and other benchmark algorithms in terms of energy-efficiency and tasks’ successful rate.

In recent years, more and more violent events are taking place in campus life. Campus bullying prevention is already the focus of current education. This paper proposes a campus bullying detecting algorithm based on surveillance video. It can actively monitor whether students are being bullied on campus. The authors use Openpose to extract bone information from video. According to the coordinate information of bone points, they extract static and dynamic features. Support vector machine (SVM) is used to classify different actions. The recognition accuracy of the classification model is 88.57%. In this way, the campus surveillance camera is able to realize real-time monitoring of bullying behavior. It is conducive to the construction of a harmonious campus environment.

With the rise of the fourth revolution of science and technology, artificial intelligence is pushed to the forefront of the world; Under the attention of the world again and again to upgrade innovation, gradually permeability; Under the attention of the world again and again to upgrade innovation, gradually into the people of various fields, including the education industry. Education is the foundation of training talents, its purpose is to improve the person’s intelligence and a kind of activity, and artificial intelligence are pretty much the same.

In recent years, led by the wave of artificial intelligence, “artificial intelligence + education” has become a very hot topic. More and more traditional educational institutions have begun to organize and layout the field of ARTIFICIAL intelligence education. Training artificial intelligence talents will become an important mission of education. Meanwhile, educational methods will change with the development of artificial intelligence, and the deep integration of artificial intelligence and education will become the development trend of the future education world. The future has come, when education and artificial intelligence meet, what kind of spark will be produced? This paper mainly discusses the application, research status and development trend of artificial intelligence in the field of education, as well as the deep integration of artificial intelligence and education.

In recent years, “artificial intelligence + education” has become a very hot topic under the guidance of the wave of artificial intelligence. More and more traditional education institutions begin to organize and lay out the field of artificial intelligence education. The cultivation of artificial intelligence talents will become an important mission of education. At the same time, the mode of education will also change with the development of artificial intelligence Integration will become the development trend of the future education world. The future has come, when education and artificial intelligence meet, what kind of sparks will collide? This paper mainly discusses the application, research status and development trend of artificial intelligence in the field of education, as well as the deep integration of artificial intelligence and education.

This paper explores the application of face recognition system in dormitory management, and designs an EXE software, which realizes the function of entering and leaving dormitory through face voucher, storing the information of students, and automatically updating relevant data. The face recognition module includes three functions: face image recognition and interception, face alignment, face feature extraction and face verification. The traditional method of HOG is used for recognition and interception, and the method of gray-scale processing and gamma normalization is used for preprocessing to reduce the influence of light. The face alignment uses the 68-point landmarks and similarity transformation to align the face. The face feature extraction uses the pre-trained FaceNet neural network to extract the 128-d feature vector of the face, which is stored in the database or compared with the face in the database to output the Euclidean distance, then compare and find the most possible person according to the distance. The database module includes two functions: storing face information and displaying all information of students MySQL is used to build the database. An automatic dormitory information management system is established through the interconnection interface between MySQL and Python.

Higher vocational education has adopted the e-learning system, and scholars have achieved a lot of results in e-learning. However, how to introduce flow ex-perience theory, extract the behavioral intention characteristics of higher vocational students, and how to integrate job requirements and skill certificates into e-learning Design and application need to be discussed in depth. We propose a UTAUT model that combines flow experience, exploring the use of behavior intention as a mediator and flow experience as the target variable. More than 7000 students from City College of Huizhou participated in the questionnaire. The Structural Equation Modeling (SEM) SmartPLS3 software was used to investigate their flow experience to use the e-learning system. The results show that perceived usefulness and facilitating conditions have an important influence on their flow experience and behavioral intentions, both have a partial mediating effect on flow experience through behavioral intention. The e-learning system of higher vocational education should promote the flow experience level of students, and strengthen the elements of employment positions and skills certificates. Suggestion: The e-learning system of higher vocational education should promote the flow experience level of students, and strengthen the elements of employment positions and skills certificates. The model of intention to use e-learning systems for senior students is innovative and effective in practice.

With the rising of the civil UAV (Unmanned Aerial Vehicle) industry and the opening of low-altitude airspace, UAVs are frequently used for privacy snooping, terrorist attacks and similar activities, which greatly harm people's safety and social security. However, in the complex urban environments, traditional low-altitude target detection methods are difficult to effectively detect small low-altitude targets which have small size and low speed. With the advent of 5G, intensive networking of 5G base station makes 5G signal is the most abundant resources of electromagnetic in the city, using 5G as external illuminator signal can not only realize the effective detection of low altitude small target, also save the cost, reduce the impact on the urban electromagnetic environment such as the radar, promoted the radar communication integration. Based on the above research background, this paper mainly studied the low altitude target detection scheme based on 5G base stations which is applicable for urban environment. This scheme using 5G base stations as radar transmitters, 5G signal as radiation source, set up the receiver receives the forward scattering signals from the target in order to achieve low altitude target detection and imaging. This paper systematically discusses the feasibility and advantages of 5G signal used as radar radiation signal, studies its radar performance, and the simulation proves its superior speed resolution and range resolution. It provides theoretical basis and support for the application of low-altitude airspace accurate detection and so on, and pushes forward the integration process of radar communication.

In order to know the development context of radar gesture recognition and predict the possible future development trends, the research and development of gesture recognition based on radar technology in recent years was sorted out. Focusing on key technologies such as dynamic gesture information perception, gesture echo signal preprocessing and feature extraction in radar gesture recognition technology, and classification algorithms for gesture recognition, the relevant literature published at home and abroad is summarized and existing methods are summarized. The performance of the system is analyzed and evaluated; the problems to be solved in the research direction are sorted out and the future research directions are prospected. The results show that radar gesture recognition technology has made great progress in human-computer interaction applications. With the deepening of related research, the gesture recognition system based on radar technology will develop towards intelligence.

As a frontier hot spot in the current image processing field, deep learning has unparalleled superiority in feature extraction. Deep learning uses deep network structure to perform layer-by-layer nonlinear transformation, which can achieve the approximation of complex functions. From low-level to high-level, the representation of features becomes more and more abstract, and the more essential the original data is described. Aiming at the problem of SAR image target detection, this paper studies the influence of the number of convolution kernels, the size of the convolution kernel and the number of convolution layers in the deep convolutional neural network on SAR target recognition.

CNN-based object detectors have been widely exploited for vision tasks. However, for specific real-time tasks (e.g. object detection on static monitor), the enormous computation cost makes it difficult to work. To reduce the computation cost for object detection on static monitor while inheriting high accuracy of CNN-based networks, this paper proposals a method with a two-stage detector using Gaussian mixture model for region proposal. We test our method on MOT16 datasets. Compared with original models, the two-stage detectors equipped with Gaussian region proposal achieve a better performance with the mAP increased by 0.20. We also design and train a light-weight detector based on our method, which is much faster and more suitable for mobile and embedded device with little drop in accuracy.

The speckle noise inherent in synthetic aperture radar (SAR) images seriously affects the visual effect of the image and brings difficulties to the subsequent parameter inversion and interpretation. However, the existing SAR image filtering methods are not effective in preserving the image edge details. In this paper, an exponential image processing (EIP) correction based lee filter denoising method is proposed to solve this problem. This method carried out a reasonable fuzzy division on the image gray histogram, and extracted its statistical characteristics from it. Such feature is used to correct the image based on the mathematical structure of EIP, and divide the filter area of the image to avoid the loss of edge information and dark details of the image. Simulation results have shown that the proposed method outperform the traditional methods in suppressing noise and protecting edge details.

In the field of artificial intelligence (AI), three-dimensional (3D) vision system is increasingly used to obtain 3D information of targets. In this paper, a method for measuring the azimuth of CCD camera’s apparent axis with high precision applied to 3D vision system is proposed. The azimuth angle of the apparent axis of CCD camera can be easily and accurately measured by using the laser projection transfer method through the horizontal two-dimensional turntable and the linear laser, which provides a method for 3D vision system calibration. This paper introduced the principle of measuring angles of the system, deduced the equation of coordinate transformation of the system, and made systematic error analysis. The results show that the measurement accuracy and reliability of this method meet the needs of 3D vision system calibration, which is much higher than the measurement accuracy and stability of geomagnetic sensors.

Overspeed has always been a very dangerous behavior for people. This may cause a variety of bad consequences such as car accidents and casualties. We need to be able to obtain the relevant information of the car while detecting the speeding now, so that subsequent punishments can be made, otherwise the perpetrators may commit the crime again. This paper proposes a high-precision, efficient method for taking photos of speeding vehicles and vehicle recognition. We directly connect the radar speed measurement module with the camera module, so that we only have one terminal for the whole system. When the radar module detects that the vehicle is speeding, it will send it directly to the camera module, so that it can capture the overspeed vehicle. This accelerates the response speed of the camera module. Therefore, when we design imaging devices, we can lower the requirements without reducing the accuracy. We can timely capture the image even if we choose the camera with low price and low quality. At last, we use image processing and support vector machines to identify the license plate. The whole system has not much equipment and can be installed in a narrow space.

In these years, meter-wave radar has gotten more and more attention from the researchers all over the world for its advantages in anti-stealth. However, the beam width of meter-wave radar is wider because of the size of radar antenna's vertical aperture, and it makes the low-angle targets detecting and tracking more difficult, which has become one of the urgent problems in radar field. In this paper, the multipath effect in low-angle target detecting will be researched, and an elevation estimation algorithm of low-angle target in meter-wave radar based on machine-learning will be proposed.

In order to register the same target in images from different sources to improve the accuracy of target recognition of multi-source images, based on the principle that the same target has the highest similarity among targets in these images, this paper proposes a new target registration algorithm by fusing features of Visible (VIS), Long Wave Infrared (LWIR) and Middle Wave Infrared (MWIR) images, which registers the same target in these images by calculating the targets similarity in different source images. Firstly, the similarity between targets in LWIR and MWIR images is calculated by using the improved structural similarity. Then, the similarity between targets in VIS and LWIR images is calculated by using Hu invariant moment feature and cosine similarity. Finally, the similarity among targets in VIS, MWIR and LWIR images is calculated by fusing these two kinds of target similarity, so that target registration of these three-source images is realized. Experimental results show that the proposed algorithm has high correct rate and accuracy of target registration. Specifically, the correct rate of target registration is 83.87% and the accuracy of target registration is higher than 0.95.

Target activity recognition has many potential applications in the fields of human-computer interaction, smart environment, smart system, etc. Recent years, due to the miniaturized design of the frequency modulated continuous wave (FMCW) radar, it has been widely utilized to realize target activity recognition in our daily life. However, the activity recognition accuracy is usually not high due to the surrounding noise and variation of the activity. To realize high accuracy activity recognition, one feasible way is to extract discriminative features from the weak radar signals reflected by the activity. Inspired by the successful application of deep learning in computer vision, in this paper, we try to explore leveraging deep learning to solve the target activity recognition task. Specifically, based on the characteristics of the FMCW signals, we design the Doppler radio images suitable for the deep network to deal with. Then, we develop a deep convolutional network to extract discriminative activity features from the Doppler radio images. Finally, we feed the features into a Softmax classifier to recognize the activity. We carry out extensive experiments on a 77 GHz FMCW radar testbed. The experimental results show the excellent target activity recognition performance.

In order to further improve the positioning accuracy of indoor positioning based on visible light communication, this paper proposes an RSS fingerprint database location algorithm based on Lambert optimization model, which significantly improves the positioning accuracy. The Internet of Things technology is used to realize the connection between devices and mobile phones, and the positioning information is transmitted to mobile phone clients in real time. To test this system, a visible light indoor positioning system based on STM32F407 development platform was built, which verified that the positioning error of the system was basically stable within 30 mm in the area of 800 mm * 800 mm. This system has stable signal transmission, high precision, small time consuming and low cost of power consumption, which has a good development and application value.

Target detection is one of the hotspots of image processing research. In the image, due to factors such as distance or light, it will affect the target detection result and increase the error detection rate. Moreover, the existing network training time is too long to meet the actual needs. In order to reduce the lack of light or shadow interference and other factors, based on the Faster R-CNN framework, this paper innovatively proposes a method to improve its feature network ResNet-101 to extract deep features of images.In order to shorten the running time, this paper introduces the region number adjustment layer to adaptively adjust the number of candidate regions selected by RPN during the training process. This paper conducts experiments on the PASCAL VOC data set. The experimental results show that the improved feature network model proposed has an accuracy improvement of 2% compared with the original feature network model. The results show that the target detection algorithm proposed in this paper has higher recognition accuracy than the original algorithm.

Through theoretical analysis, the calculation method of dual basis multiplication in GF (28) field based on CCSDS Berlekamp is given. Based on this calculation method, a VLSI architecture for parallel multiplication and serial operation in circuits is proposed. At the same time, the hardware resource occupation and the timing performance of each VLSI architecture are analyzed in detail.

Terahertz three-dimensional imaging system can realize the detection and imaging of near-field targets with high frame rate and high resolution, and can provide more comprehensive information about the three-dimensional geometric distribution structure of the target and the imaging scene. It is suitable for the current high real-time requirements Security inspection, seeker terminal guidance, military reconnaissance and other fields. The high-resolution three-dimensional imaging technology of radar targets in the terahertz band is of great significance to the development of radar technology and the application of radar imaging. In this paper, the research background and significance of the terahertz near-field imaging technology, radar three-dimensional imaging technology, the development status of terahertz radar system and terahertz radar imaging algorithm are reviewed, and the existing problems of terahertz near-field imaging technology are summarized and prospected.

Various path or route solving algorithms have been widely researched for the last 30 years. It has been applied in many different robotic systems such as bomb sniffing robots, path exploration and search rescue operation. For instance, an autonomous robot has been used to locate and assist a person trapped in the jungle or building to exit. Today, numerous maze solving algorithms have been proposed based on the some information available regarding the maze or remotely control. In real scenario, a robot is usually placed in an unknown environment. It is required for the robot to learn the path, and exhibit a good decision making capability in order to navigate the path successfully without human’ assistance. In this project, an Artificial Intelligence (AI) based algorithm called Recursive Backtracking Depth First Search (RBDS) is proposed to explore a maze to reach a target location, and to take the shortest route back to the start position. Due to the limited energy and processing resource, a simple search tree algorithm has been proposed. The proposed algorithm has been evaluated in a robot that has the capability to keep track of the path taken while trying to calculate the optimum path by eliminating unwanted path using Cul-de-Sac technique. Experimental results have shown that the proposed algorithm can solve different mazes. The robot has also shown the capability to learn and remember the path taken, to return to the start and back to target area successfully.

FPGA emerges as a very promising AI chip and algorithm hardware accelerator. However, the FPGA is susceptible to complex and changeable environment, which leads to circuit configuration information faults. To address this issue, we propose FTEI, a fault tolerance model of FPGA with endogenous immunity. At fault detection phase, we put forward a fault detection models based on optimized logistic regression classification and use it to establish a fault model matching library. During fault recover stage, we use fault configuration library and online evolution to recover faults. In order to improve the success ratio of online evolution, we propose RLAGA, an adaptive genetic algorithm based on reinforcement learning. Experiments on typical functional circuits, 8-bit parity verifier and 2-bit multiplier, demonstrate that the fault detection accuracy rates reach 94.4% and 93.2%, and the fault recover success rates of RLAGA are 100% and 90%, which significantly improves FPGA errors detection and recover effectiveness.

Now the monitoring equipment such as cameras has been widely used in social life. In order to solve the problem that the current monitoring equipment relies on manual screening for the recognition of abnormal human action and is not time-efficient and automatic, an intelligent monitoring system for human action is designed in this paper. The system uses object detection, classification and interactive recognition algorithm in deep learning, combines 3D coordinate system transformation and attention mechanism model. It can recognize the local human hand actions, head pose and a variety of global human interaction actions in the current environment in real time and automatically, and judge whether they are abnormal or special actions. The system has high accuracy and high speed, and has been tested successfully in laboratory environment with good effect. It can also reduce labor costs, improve the efficiency of security monitoring, and provide help for solving urban security issues.

The data generated in the process of planning, construction, and management of construction and building is diverse and large in scale, and there is an urgent need for a unique, consistent, and efficient code. Beidou grid position coding stipulates its grid selection and coding rules, as well as spatial position information identification, transmission and big data processing, which can have good scalability in the field of building spatial identification position coding. Based on the coding rules of the Beidou grid position code, this paper proposes a construction building space identification position coding technology to code the construction building. The coding is unique, consistent and efficient. The application of coding rules is introduced in the paper, including code generation and coding index, as well as the corresponding query algorithm, which is verified by engineering experiments.

Generally, the output power of photovoltaic (PV) inverter will match the load requirement. And at the beginning of the design the load power is less than the maximum output power of PV cells to ensure the system operation stable when the PV inverter operates in islanded mode. However, it causes the energy waste of PV cells. Therefore, more and more PV cells are combined with other energy sources to form the microgrid system in order to reasonably plan the power output of each energy source. Droop control is usually used to achieve the power distribution of parallel inverter in microgrid system. However, the traditional methods of adjusting the droop coefficients or adding virtual impedance cannot automatically achieve the maximum utilization of output energy of PV cells. Thus, a novel droop control method has been proposed to achieve the maximum power output of PV (MPO-PV) unit in this paper, where the PV units of parallel system always operate at the maximum power and the other inverters make up the remaining power required by the load, with effective improvement of the utilization rate of renewable energy sources (RESs). Meanwhile, the control parameters of the improved droop loop have been designed by the small signal modeling and system stability analysis. Finally, the validity of the proposed method has been verified by experimental results.

Sensitive word recognition technology is of great significance to the protection of enterprise privacy data. In electric power custom services systems, the dialogue texts recording the conversational information between electric power customers and the customer services staffs contain some sensitive information of electric power customers. However, the colloquialism and synonyms in dialogue texts often make sensitive information recognition more difficult. In this paper, we proposed an out-of-vocabulary (OOV) approach for recognizing sensitive words in the dialogue texts of electric power customer services. We combine the semantic similarity based on word embeddings and structural semantic similarity based on HowNet for recognizing sensitive OOV words in the dialogue texts. The related experiments were made, and the experimental results show that our method has higher recognition accuracy in comparison with the popular approaches.