Artificial Intelligence in China

Inhaltsverzeichnis

1. Frontmatter

2. Inception Based Medical Image Registration

   Wenrui Yan, Baoju Zhang, Cuiping Zhang, Jin Zhang, Chuyi Chen

   Abstract

   Biological tissue has strong absorbability and strong scattering, which may lead to edge blur, low signal-to-noise ratio and so on, which affects the registration of medical images. This chapter simulates and makes the biological tissue imitation image data set under the above conditions, cleverly designs a regression structure embedded in the U-net Inception module, and forms an unsupervised deep learning medical image registration method. It is applied to the biological tissue imitation data set for the first time. The experimental results show that the image registration method based on SIFT, ORB, BRISK, AKAZE and so on can not successfully complete the registration work because it can not find a sufficient number of effective key points on the data set, SURF although it can complete the registration work. But the effect is weaker than the proposed method. Based on the medical image registration method proposed by Inception, this chapter can effectively solve the problem of biological tissue image registration with strong absorption and strong scattering.

3. Research on Smart Home System Based on Internet of Things

   Hai Wang, Peng Sun, Guiling Sun, Zhihong Wang, Xiaomei Jiang, Chaoran Bi, Ying Zhang

   Abstract

   In the era of Internet of Things (IOT), a more convenient and efficient life concept has created by the smart home system. In this integrated home solution, in order to realize the intelligent control, information technique is frequently utilized to interconnect ordinary household appliances to form an internal family network. With the help of sand table, the real environment is simulated in this design. Different types of sensors are integrated on ZigBee module which controller is CC2531. The intelligent gateway is responsible for analyzing and processing the environmental information obtained by the sensors. The proprietor can control one or more household electrical appliance and set personalized and intelligent home control scenarios according to the actual needs, such as home mode, out mode, sleep mode, etc.

4. Autoencoder-Based Baseline Parameterized by Central Limit Theorem for ICS Cybersecurity

   Gang Yue, Zhuo Sun, Jianwei Tian, Hongyu Zhu, Bo Zhang

   Abstract

   Industrial control system as the core of the industry is concerned about the cybersecurity problem and vulnerable to be threatened by the cyber-attacker. However, the conventional IDS aims to mine intrusion features and realizes intrusion detection by matching the abstract features of intrusion, so it could not recognize unknown and zero-day intrusion. In fact, the ICS as the closed-loop control system is different from the commercial internet and has stable interactive features. In the paper, we analyze the ICS network interaction and construct a parameterized baseline by an autoencoder to detect the intrusion. The experiment with an open ICS dataset shows that this baseline could achieve intrusion detection accuracy above 90% and the false alarm rate below 5%.

5. Secrecy Capacity-Approaching Neural Communications for Gaussian Wiretap Channel

   Jingjing Li, Zhuo Sun, Jinpo Fan, Hongyu Zhu

   Abstract

   Recently, some researches are devoted to the topic of end-to-end learning a physical layer secure communication system based on autoencoder under Gaussian wiretap channel. However, in those works, the reliability and security of the encoder model were learned through necessary decoding outputs of not only legitimate receiver but also the eavesdropper. In fact, the assumption of known eavesdropper’s decoder or its output is not practical. To address this issue, in this paper we propose a dual mutual information neural estimation (MINE) based neural secure communications model. The security constraints of this method is constructed only with the input and output signal samples of the legal and eavesdropper channels and benefit that training the encoder is completely independent of the decoder. Moreover, since the design of secure coding does not rely on the eavesdropper’s decoding results, the security performance would not be affected by the eavesdropper’s decoding means. Numerical results show that the performance of our model is guaranteed whether the eavesdropper learns the decoder himself or uses the legal decoder.

6. Image Compression Based on Mixed Matrix Decomposition of NMF and SVD

   Zhiyang Zhao, Baoju Zhang, Cuiping Zhang

   Abstract

   Image compression has always been a key research hotspots in image processing. An efficient image compression approach will not only save considerable storage resources, but also exceedingly ease the communication pressure of the network transmission, which has great research significance and practical value. Since the essence of the image is matrix, so mixed matrix decomposition of NMF and SVD is introduced to perform two-level compression framework on images. The experimental results demenstrated that this approach based on mixed matrix decomposition had a CR with larger dynamic range through flexible parameter adjustment and the PSNR of the restored image is 29 dB–36 dB. It verifiy that this method is effective.

7. MocNet: Less Motion Artifacts, More Clean MRI

   Bin Zhao, Shuxue Ding, Mengran Wu, Guohua Liu, Chen Cao, Song Jin, Zhiyang Liu, Hong Wu

   Abstract

   Magnetic Resonance Imaging is a common way of diagnosing related diseases. However, the magnetic resonance images are easily defected by motion artifacts in their acquisition process, which severely affects the clinicians’ diagnosis. To resolve the problem, we propose a motion correction network (MocNet) to correct motion artifacts. The experiments of motion artifacts simulation demonstrate that our MocNet outperforms the comparison methods with a mean PSNR of 34.397 ± 3.155 dB and a mean SSIM of 0.971 ± 0.015.

8. The Application Exploration of Digital Twin in the Space Launch Site

   Cai Hongwei, Zhou Bo, Yang Hui, Li Xu

   Abstract

   In view of the future development of space launch site, combined with the present situation and development of digital twin technology, the paper puts forward the idea of the construction of digital twin in the space launch site. The system analyzes the application prospect of digital twin technology in the space launch site, and designs the architecture of the digital twin of the space launch site, which is concluded that the digital twin body of the space launch site will completely change the operation management mode, the equipment guarantee mode, the task process, the organization command, and so on.

9. Reverse Attention U-Net for Brain Grey Matter Nuclei Segmentation

   Mengran Wu, Bin Zhao, Chao Chai, Guohua Liu, Zhiyang Liu

   Abstract

   It has been reported that abnormal high iron deposition in brain grey matter nuclei is closely related to neurodegenerative diseases. Therefore, precise nuclei segmentation is beneficial to further explore the pathological mechanism of neurodegenerative diseases. However, segmenting nuclei manually is an extremely time-consuming work. To this end, we proposed a Reverse Attention U-Net (RAU-Net) for segmenting nuclei automatically, where multiple reverse attention (RA) modules are added between the encoder and decoder to aggregate different features. In the meanwhile, we use the estimations of segmentation maps at multiple levels as the guide information for the RA module. The model with RA module implicitly erases the predicted nuclei regions while highlighting background, which guides the network to explore the missing nuclei parts sequentially. Experimental results on our nuclei dataset imply that the RAU-Net performs favorably against the state-of-the-art methods.

10. Transmit RRH Selection in User-Centric Cell-Free Massive MIMO Using Discrete Particle Swarm Optimization

    Jiaxiang Li, Yingxin Zhao, Hong Wu, Zhiyang Liu

    Abstract

    Cell-free massive multiple input multiple output (MIMO) has been expected to improve the spectrum efficiency in future mobile communication systems. To reduce the fronthaul link burden, a user-centric structure is expected, where a subset of remote radio heads (RRHs) are expected to form a virtual cell and serve the user. However, how to select serving RRHs for each user so as to maximize the rate performance remains to be an open question. In this paper, by assuming that each RRH can only be selected by at most one user, an improved discrete particle swarm optimization (IDPSO) has been proposed to accelerate its convergency. To further reduce the fronthaul link burden, a large-scale-fading-based fitness function is also adopted. Numerical results reveal that the proposed algorithm is able to achieve a good tradeoff between sum rate performance and the computational complexity.

11. Research on Emotion Recognition Based on GA-BP-Adaboost Algorithm

    Ruijuan Chen, Xiaofei Diao, Zhihui Sun, Guanghua Deng, Zhe Zhao, Zhuanping Qin

    Abstract

    With the development of wearable portable electrocardio (ECG) acquisition devices, real-time emotion recognition based on ECG signals becomes possible. In order to study the difference in the characteristics of human heart rate variability in different emotional states, and classify the positive and negative valence emotions, the emotional induction paradigm was designed, and a wearable ECG collection device was used to collect the ECG signals in the corresponding emotional state. Using heart rate variability as a characteristic parameter, an emotion recognition model based on GA-BP-adaboost algorithm is proposed. This model uses genetic algorithm to optimize the threshold and bias of BP neural network, and uses adaboost ensemble learning method to construct multiple BP neural networks as strong classifier. Results show that the GA-BP-adaboost algorithm has a correct recognition rate of 84.27% for positive and negative emotion samples. The algorithm proposed in this paper can significantly improve the accuracy of classification, which is of great significance in the research of emotion recognition based on heart rate variability.

12. Information Extraction of Air-Traffic Control Instructions via Pre-trained Models

    Xuan Wang, Hui Ding, Qiucheng Xu, Zeyuan Liu

    Abstract

    The air traffic controllers always use air-traffic control instructions (ATC instructions) to command aircrafts. The ATC instruction consists of some situation mentions such as flight number, status, and target location etc. The deep-learning based approach can extract such information for situation awareness. In practice, it is difficult to prepare huge amount of labelled ATC instructions for training the deep-learning model due to expensive costs of handcraft annotations. The large scale pre-trained model (PTMs) can solve this problem by “pre-training” and “fine-tuning”. This paper proposes: 1) pre-trained models to extract information from few scale ATC instructions; 2) the probing task to find which layer of model achieves the best performance of information extraction task.

13. Medical Image Segmentation Using Transformer

    Qian Wang, Longyan Li, Bo Ni, Yu Li, Dejin Kong, Chen Wang, Zan Li

    Abstract

    For the past few years, the U-Net structure shows strong performance in the field of medical image segmentation. However, due to the inherent locality of convolution operations, U-shaped structures are often limited in modeling long-range dependencies. Transformer, a global self-attention mechanism designed for sequence-to-sequence prediction, has been successfully used in the field of computer vision. In this paper, we propose a novel network, named TransHarDNet. HarDNet, which is a low memory traffic CNN. We combine it as backbone with Transformer. Our network enables the global semantic context information and low-level spatial details of the input image to be captured more effectively. We evaluate the effectiveness of the proposed network on five medical image datasets.

14. Satellite Online Scheduling Algorithm Based on Proximal Policy

    XueFei Li, Jia Chen, XianTao Cai, NingBo Cui

    Abstract

    Aiming at the problems of high labor cost, low task execution efficiency, and unable to adapt to the normalized dynamic observation requirements caused by satellite scheduling under off-line control of satellite and ground, this paper considers each element of the satellite online scheduling process, and according to the Markov nature of satellite online scheduling process, combined with the proximal policy optimization algorithm, proposes a satellite online scheduling algorithm, which effectively solves the problems of satellite resource conflict and time window conflict in the scheduling process, and realizes the real-time response to the task.

15. Generation Method of Control Strategy for Aircrafts Based on Hierarchical Reinforcement Learning

    Zeyuan Liu, Qiucheng Xu, Yanyang Shi, Ke Xu, Qingqing Tan

    Abstract

    With the increasing density of air traffic and the complexity of the terminal sector, air traffic controllers will face more challenges and pressures in ensuring the safe and efficient operation of air traffic. In this work, an artificial intelligence (AI) agent is built to handle dense, complex and dynamic air traffic in the future. In this work, an artificial intelligence (AI) agent based on deep reinforcement learning is built to mimic air traffic controllers, such that the dense, complex and dynamic air traffic flows in terminal airspace can be handled sequentially and separated. To solve the problem, hierarchical reinforcement learning method is proposed, the flights choose agent and the flights action agent are achieved by DDQN. Results show that the built AI agent can guide 16 aircrafts safely and efficiently through Sector 01 of Nanjing Terminal, simultaneously.

16. Finding Significant Influencing Factors of Core Quality and Ability Development of Teachers Based on Improved Genetic Algorithm

    Jian Dang, Yueyuan Kang, Xiu Zhang, Xin Zhang

    Abstract

    The professional development of teachers is crucial for the education development of a country. For understanding the core qualities and ability development of teachers, it is very important to identify the associated influencing factors. In this paper, an improved genetic algorithm is used to find out the significant influencing factors among the many influencing factors of teachers’ core qualities and ability development, the idea of the improvement is to introduce elite solution retention strategy and adaptive mutation probability on the basis of traditional genetic algorithm. At the same time, the collected samples are processed by back propagation neural network, and the accuracy of its prediction is taken as the criterion to measure the significant degree of the influencing factors. The experimental results show that the improved genetic algorithm is more effective than traditional genetic algorithm in finding the significant influencing factors.

17. A New Augmented Method for Processing Video Datasets Based on Deep Neural Network

    Wei Wang, Haiyan Wang, Fuchuan Ni

    Abstract

    Large datasets are required for deep learning to achieve good performance. However, there is a lack of sufficient training datasets in many research fields, which may become a shortcoming of computer vision applications. This article provided a new data augmentation method for making training small datasets, which could be divided into two steps: 1. Unbalanced sampling based on information density. 2. Splicing images to form a dataset. Different information density dataset combinations had been used for testing the model generalization. The enhanced loss function which consisted of label smoothing loss and cross-entropy loss had been used to minimize the model preference during training models. Finally, with the same amount of data, the Mean Absolute Error (MAE) of the model with our sampling method could get 55% increase compared with the traditional sampling method. The best MAE could reach 0.98 if the splicing method had been adopted. The results showed that this augmented method was suitable for scenarios with small sample size, especially video datasets. To get the best performance, the splicing method was a nice choice to optimal model generalization performance.

18. Small-Object Detection with Super Resolution Embedding

    Wenyi Tang, Qiucheng Xu, Hui Ding, Lianghao Wu, Yanyang Shi

    Abstract

    The detection performance of ground vehicles on the airport surface in video-sensor based air traffic control surveillance has not been satisfactory compared to larger planes, especially in panoramic images of remote tower applications. Inspired by the success of GAN-based super resolution and oversampling augmentation methods, we apply a new super-resolution network to resize cropped small samples and augment each of those images by resize-and-pasting small objects many times. It allows us to trade off the quantity of the detector on large objects with that on small objects. We propose an architecture with two components: ESRGAN and Detection network. We use residual dense blocks for ESRGAN, and for the detector network, we use one state-of-the-art detector (YOLOv3). Extensive experiments on a public (car overhead with context) dataset and another self-assembled airport surface dataset show superior performance of our method compared to the standalone state-of-the-art object detectors.