针对高校计算机网络课程思政的教学实际,分析新工科背景下课程思政要解决的3个关键问题,提出"理论—方法—实践"3层建设框架,介绍计算机网络课程思政建设的具体实践方法,最后利用问卷调查法,通过最近学期的实际教学数据汇总,基于主观评价方法对课程思政的实践效果进行分析与评价.

Document layout analysis (DLA), aiming to extract and classify the structural regions, is a rather challenging and critical step for many downstream document understanding tasks. Although the fusion of text (semantics) and image (vision) features has shown significant advantages for DLA, existing methods either require simultaneous text-image pair inputs, which is not applicable when only document images are available, or have to resort to an optical character recognition (OCR) preprocessing. This paper learns the visual-sematic representation for DLA only from the imaging modality of documents, which greatly extends the applicability of DLA to practical applications. Our method consists of three phases. Firstly, we train a text feature extractor (TFE) for document images via cross-modal supervision that enforces the coherence between the outputs of TFE and the text embedding map generated by Sent2Vec. Then the pretrained TFE gets further adapted using only the document images and extracts shallow semantic features that will be further fed into the third stage. Finally, a two-stream network is employed to extract the deep semantic and visual features, and their fusion is used as the input to a detector module, e.g., the RPN (Region Proposal Network), to generate the final results. On benchmark datasets, we demonstrate that the proposed TFE model outperforms main-stream semantic embedding counterparts and that our approach achieves superior DLA performance to baseline methods.

In this paper, we present a nonparametric Bayesian dictionary learning method for hyperspectral image (HSI) denoising, which exploits two intrinsic properties of HSIs: the spectral correlation and the spatial nonlocal similarity. We first extract full band patches from the contaminated HSI and then cluster them into groups. Then we represent each group as a fourth order tensor so that the spatial-spectral-nonlocal features can be preserved via Tucker decomposition. As tensor rank estimation is still an open problem, we propose a hierarchical nonparametric Bayesian optimization scheme for the factor matrices inference, in which Beta-Bernoulli process priors are exerted on the core tensor to allow for the datadriven rank determination over different tensors. Furthermore, we optimize the tensor decomposition result via the Nonnegative least square method (NNLS) to constrain the positivity of dictionary atoms. Numerical experimental results verify the efficacy of our method compared with state-of-the-art HSI denoising methods.

Aerial surveillance information fusion is a key subsystem in air traffic control system. An excellent aerial surveillance information fusion algorithm can get more accurate position estimation on the aircraft. The commonly used algorithm for aerial surveillance information fusion is Kalman filter. The filtering accuracy of the Kalman filter algorithm is affected by accuracy of its parameters. When parameters are inaccurate, it may even cause the filter to diverge, so how to determine the parameters of Kalman filter is a key problem. This paper proposes a filtering model that integrates Back Propagation Neural Network, Generalized Regression Neural Network and Kalman filter. The parameters of Kalman filter are adjusted during filtering process dynamically by neural networks, so that the adaptability of traditional Kalman filter is enhanced. The actual radar measurement data is used for filtering experiments, and experimental results show effectiveness of this model.

The most commonly used algorithm for aviation surveillance is Kalman filter. The accuracy of Kalman filter is affected by the accuracy of its parameters. When the parameters may change with environment change, the accuracy of traditional linear Kalman filter will be affected, in severe cases, filtering divergence will occur. This paper proposes an aviation surveillance filtering model that treats Kalman filter as the kernel of Recurrent Neural Network, uses Back Propagation Neural Network to predict parameters of Kalman filter. Which let Kalman filter be trainable and have ability to estimate parameters dynamically. Moreover, actual radar measurement data is used for radar filtering experiment, and the experiment results show the feasibility of this model, and show that this model has better accuracy and adaptability than traditional Kalman filter.

With the development of BlockChain [1], more research attention is placed on Byzantine Fault Tolerance. This paper proposes Scalable Efficient BFT protocol (SeBFT) based on MinZyzzyva [4]. Replicas in SeBFT can join or exit freely in groups. SeBFT also inherits MinZyzzyva's excellent performance with the communication complexity is O(n) and replicas number required within group is 2f+1.Moreover, this paper also proposes HSeBFT which applies SeBFT to hierarchical network and further expands the maximum network scale and provides a BFT solution for multi-level network. Compared with the existing scalable solution, SeBFT has great advantages in algorithm complexity, delay, scalability and fault tolerance.

With the great success of cryptocurrencies, people have paid great attention to blockchain technology and Byzantine fault tolerance (BFT) for consensus. Practical Byzantine Fault Tolerance protocol (PBFT) proposed by Castro and Liskov in 1999 is recognized as the most classic protocol for solving the Byzantine generals problem, but it still has many fatal disadvantages. Firstly, its communication complexity in the consensus process has reached O(n 2 ), so that it can only be applied in a small-scale system. Secondly, it works in a completely enclosed environment, nodes are not allowed to participate or exit the network without restarting the system, making it difficult to get applied in the actual system. In order to solve these problem, this paper proposes the High Performance and Scalable Byzantine Fault Tolerance, an improved BFT protocol, called HSBFT. In normal case operation, the HSBFT’s communication complexity is reduced to O(n), and it also has the ability to be scalable, making it adaptable to more diverse practical scenarios.

Aerial surveillance information fusion is a critical step in air traffic control systems. The surveillance system employs fusion algorithms to process measurement data from surveillance source such as radar, reduce measurement error, predict the future position of the aircraft, and provide the supervisor with predicted data according to the time interval set by the system. The most widely used fusion algorithm in air traffic control systems is the Kalman filter algorithm. This algorithm has been applied for several decades in related fields, but it requires considerable manpower and time to adjusting parameters, so it's use-cost is relatively high. The Kalman filter algorithm needs to adjust the parameters of the algorithm in different scenarios according to the position-accurate calibration data, considering the sector of the radar, the distance between the aircraft and the surveillance source, and the terrain. The process is very similar to the training process of supervised learning. This paper attempts to use neural networks to fuse surveillance information in order to replace the complex manual tuning process. Firstly, the neural network has the ability to represent complex features of things. The appropriate number of network layers and neurons can characterize the distribution of measurement errors in surveillance data. Secondly, we could train the network by using sufficient surveillance data and accurate target data to automatically capture features. It reduces labor and time commitments, and reduce the use-cost of the algorithm. This paper analyzed the shortcomings of the traditional fusion algorithm, introduced the overall structure of the fusion system, focused on the application of the neural network algorithm in the system, and gave the fusion result data of the algorithm.

Blockchain is an emerging decentralized infrastructure and distributed computing paradigm. Blockchain has been widely used in many industries including finance, energy, and cloud computing. Consortium blockchain is known as the permissioned blockchain, in which each user needs to be authorized. Distributed consensus is the core issue of consortium blockchain technology. In order to solve the disadvantages of using Proof of Work (POW) algorithm in consortium blockchain, such as computing resources waste, long block confirmation delay and low throughput, we designed and implemented a high performance blockchain consensus algorithm based on message passing technology called HPBC. HPBC algorithm can work well in asynchronous networks with Byzantine nodes. Results from theoretical analysis and performance evaluation show that HPBC algorithm provides safe and reliable distributed consensus services in consortium blockchain with low overhead. Finally, we summarized HPBC algorithm and briefly explore its further research directions.

On the basis of Time Domain Analysis, this paper proposes a method for music beat tracking. Through this method, music beats are detected and tracked by timestamp and intensity value. Generally, the beat-areas of music signal converge more energy than other areas, therefore, the spots of beat can be filtered out by a tracking algorithm with a dynamic threshold value. In this paper, dynamic threshold value in tracking algorithm is modified by using two sliding windows, which are Prediction Window and Detection Window. Also, a new indicator which indicates the stationarity of the signal is proposed. This factor can distinguishes the music signal with rhythmic beats from which with lone-tone and noise in time-domain. The experimental result proves the simplicity, adaptability, and robustness of this method, and it is an efficient algorithm on music beat tracking.

Based on the theories of frequency domain and time domain signal processing, wavelet analysis, and singular value decomposition (SVD), an effective method for content based music feature extraction is proposed in this paper. Music feature can be divided into three parts by this method, which are frequency feature, auditory perceptual feature, and statistical characteristic of beat. The characteristic of each music can be well described by these features. The results of logistic regression classification model and linear support vector machine (SVM) classification model which is on a data set consists of several different styles of music and use the feature extraction method in this paper show the high precision of 95.33% in average, and also prove the effectiveness of the proposed method. Feature extraction is the foundation of content based recommendation, retrieval, classification, and cluster. Hence this method has good prospect in these area.

With the development of mobile devices,face recognition technology has been widely used in the mobile terminal.Changes are constantly taking place in the backgrounds and illumination conditions of images collected by mobile phone,and the accuracy of face recognition is affected.This paper presents a face recognition system based on BP neural network,and pre-process the images by histogram equalization,smoothing and el-liptical mask so as to improve the accuracy of face recognition.In this paper,a real face recognition system is built and the effectiveness of the algorithms is verified.

It is difficult to get the original information from a MD5 hash since MD5 is an irreversible hash algorithm. MD5 decryption is based on MD5 calculation for brute force attack, which requires great computing resources. This paper presents an approach for MD5 calculation and decryption on GPU, which has high concurrency. The CUDA program performed on a PC with NVIDIA GTX 560TI graphics card. The experimental result has shown that the calculating speed is 150 million words per second, increasing from 10 to 20 times compared to program run on CPU (Core i7-950 @3.07 GHz). GPU has great potential in future general-purpose computing and concurrent computing.

Collaborative filtering is one of the most widely used techniques in building recommender systems for individual users, on the other hand, many group recommenders have been build due to the fact that many daily activities are group activities. In this paper, we study the technique which collaborative filtering used to predict users' preference for unknown items base on their neighbors' preference; and the various preference aggregation strategies used in prior group recommendation systems. The goal of this paper is to examine the relationship between various aggregation strategies and the size of the group in terms of maximizing the global satisfaction of the group.

Recommemder System is becoming more and more important for getting information in recent 20 years. But recommender system has the weakness of extreed large scale that makes it delayable for recommendation, which making it cannot offer real-time service. Business recommender system is general divided into two parts, the on-line recommend part and the off-line calculation part. It precomputes the off-line part to get quicker recommendation when needed. Pretended real-time recommendation is a compromise with the growing and changing system. We propose the better way to get better real-time service by processing the off-line calculation on GPU , which is a high-speed parallel processor , to speed up the first part of recommender system to get more real-time service. Our experiments show, the off-line part can speed up 19 times when using GPU, and the larger of the data scale, the better it can improve.

LDA (Latent Dirichlet Allocation) is a text modeling algorithm based on a generative probabilistic model. It is widely used to discover latent topics among a set of documents. Mahout has implemented LDA algorithm, however, the execution time of the LDA program is very long when processing a large amount of documents, because the documents are processed in sequence. This paper introduces a method to modify this program with CUDA toolkit provided by NVIDIA, in order that a group of documents could be processed in parallel on GPU. Using this method, the LDA program could be accelerated greatly.

Document clustering is one of the most important tasks in text mining. In clustering algorithms, high-dimensional vector is usually used to represent a document which causes that the algorithms are often computationally expensive. On the other hand, Graphic Processing Unit (GPU) is increasingly important in parallel computing due to its powerful parallel capacity and high bandwidth. This paper implements a GPU-based Harmony K-means Algorithm (HKA) with NVIDIA's Compute Unified Device Architecture (CUDA), and uses it for document clustering. In our experiment, our GPU-based program can acquire a maximum 20 times speedup in contrast with CPU-based program.

The traditional algorithm of BP neural network based on Map Reduce framework has been researched.But the convergence speed of traditional BP algorithm is quite slow.The training is easily into local minimum spot.The iterations of processing training data are quite frequency,a great waste of resources.The proposed design thus parallels the improved BP neural network based on Map Reduce.The adjustment of self-adaptive learning rate and momentum factors are implemented in this design.The proposed design also uses maximum classification probability factor and preprocess method to enhance the correctness.Experimental result shows that improved parallel algorithm costs 17 times shorter than the primary.The accuracy is also improved further.

Generating shorter testing sequence is an important issue in protocol conformance test.This paper proposes a method for generating BUIO sequence by using UIO sequence.Some UIO sequences switching to BUIO sequences can bring the cost reduce of the BUIO generation.The heuristic sequences generation algorithom based on UIO and BUIO is analyzed and improved to ensure the automatic generation of test sequence.Test sequence is obtained through applying the algorithm to protocol ECMA-203,which is compared to the test sequence generated by using Rural Chinese Postman algorithm and UIO sequences.

Mobility support technology for both traditional internet and mobile internet encounters a great deal of challenge,one important factor among them is that the performances of current transport layer protocols,especially reliable transport protocols,play poorly in mobile and wireless environment.This paper focuses on the study of Stream Control Transmission Protocol(SCTP) and its mobility support function,the SCTP performance optimization schemes in Mobile IP networks is analyzed as well.