Abstract:RDMA networks have the characteristics of high bandwidth, low latency, and low CPU load, and are used in a wide range of data-intensive tasks, such as deep learning, high-performance computing, and data analysis. The imple-mentation of RDMA requires software and hardware support. In a cloud environment, the RDMA virtualization solutions help multiple users to share the high performance of RDMA network, while achieving unified management of RDMA networks. In this study, solutions of RDMA virtualization in recent years are collected, which cover virtual ma-chines and container environments. Then, these solutions are classified and compared. Finally, the existing problems and future development in RDMA virtualization are analyzed.

Abstract:Text classification refers to the process of letting a computer learn to complete the classification of content by some classification algorithm under the classification system of text. Algorithms related to text classification have been applied to web classification, digital libraries, news recommendation, and other fields. Based on the characteristics of short text classification tasks, this study proposes a hybrid short text classical model based on multi-neural networks. By reconstructing the text features of the keywords extracted from the short text content, and using the vector fusion as the input of the multi-neural network model, the characteristics of the FastText model and the TextCNN model are taken into account. The experimental results show that compared with the current popular text classification algorithms, the multi-neural network hybrid short text classification model shows more superior algorithm performance on multiple indicators such as accuracy, recall, and F1 score.

Abstract:Person re-identification (Re-ID) has attract lots of attention in computer vision, which is of great significance to the development of intelligent security and video surveillance. Currently, most existing methods focus on the person re-identification based on visible light, and have achieved good performance. However, the visible light camera cannot be used normally in the dark night, and the new generation of cameras can automatically switch the mode between infrared and visible settings for 24 hours monitoring. Therefore, some scholars have started to study the RGB-IR cross-modality pedestrian re-identification. This paper introduces the Re-ID and cross-modality Re-ID respectively from the definition, research difficulties, and development status. For RGB-IR cross-modality Re-ID, according to the types of methods, they are divided into three categories: methods based on unified feature models; methods based on metric learning; and methods based on modal transformation. We also describe the corresponding datasets and evaluation protocol. Besides, we analyze and summarize the performance of existing algorithms. Finally, the future development directions of RGB-IR cross-modality Re-ID are summarized.

Abstract:Outgoing Long-wave Radiation (OLR) is an important parameter in the study of precursor anomalies. OLR data contains important precursor information. Therefore, this study proposes a pre-seismic anomaly analysis method based on OLR data, the so called spatial-temporal analysis method. Then we apply the proposed method to analyze the Wenchuan earthquake on May 12, 2008, the Japan earthquake on March 11, 2011, and the Lushan earthquake on April 20, 2013. The experimental results show that there is important precursor information in OLR data. About three months before the occurrence of the three earthquake cases, the abnormal values in the epicenter area and its surrounding areas increased significantly, and the abnormal phenomenon in the surrounding area might be more obvious than the epicenter area. In addition, it is also found that before the occurrence of the three earthquakes, there were obvious peaks in the north direction anomaly curves in the studied areas. This finding is instructive for future work. Whether this rule is universal needs further verification.

Abstract:Aiming at the problem that the attribute value is a probabilistic hesitant fuzzy number and the decision maker’s attitude to risk is different, a probabilistic fuzzy multi-attribute decision-making method considering risk preference is proposed. First, considering the decision maker’s hesitation may affect the decision-making effect, a hesitation formula expressed by the difference between the number of elements in the probability hesitant fuzzy element is given. The extended Hamming distance and the extended Euclidean distance are defined based on the hesitation and the difference of element values. Then, a foreground decision matrix is established based on the expected values given by decision makers, and the maximum weight method is used to calculate the attribute weights. Based on this, the comprehensive foreground value of each plan is calculated and ranked. At last, the example analysis of purchasing ERP system software verifies the validity and rationality of the proposed method.

Abstract:Aiming at the time-consuming, laborious, and weak expansibility of user intention recognition in question answering robots based on template matching, keyword co-occurrence or artificial feature set, this study proposes a model based on the combination of grid memory network (LSTM+CRF+Lattice) and Convolutional Neural Network (CNN) combined with the characteristics of geological literature question answering. In this hybrid model, users’ query intention recognition is regarded as a classification problem. Firstly, the grid memory network is used to identify the named entity and extract the relationship of the text information, then the CNN is used to classify the attributes of other text information input by users, and then the classification results are transformed into a structured way to meet the query of knowledge graph, and finally realizes the attribute mapping of user intention recognition. Experiments show that it is very helpful to improve the accuracy rate when considering the characteristics of geological knowledge.

Abstract:Nowadays, the Internet recommendation system has become a hot topic. Automatic recommendation has greatly facilitated people’s life and helped people find the most interesting key information from the massive information. Now news information is generated every moment on the Internet, and the existing information is a very large data set, which can help to count the user preferences and popularity of news content. At present, there are many kinds of recommendation systems on the Internet. They comprehensively consider the characteristics of users and articles to be recommended. Based on the data on various social media on the Internet, they build models and can use these models for accurate personalized user recommendation. The existing recommendation system is usually a supervised learning system which takes a lot of user characteristics into account. These methods often ignore the following issue: the recommendation strategy in the history is often imbalance. Through the existing historical records, we cannot guarantee an unbiased result. So in this study, we propose a kind of personalized recommendation based on counterfactual learning. This method has stronger theoretical guarantee and also shows better algorithm performance than existing methods in the experimental results.

Abstract:Small and medium-sized enterprises (SMEs) are the important pillars to support China’s economic development, and high-performance computing plays an important role in promoting scientific and technological innovation. The combination of these two part is of great significance. However, the high threshold of high-performance computing technology hinders the effective application of this technology by SMEs to a certain extent. Based on the high-performance computing environment, we build a convenient and easy-to-use high-performance computing application community. The purpose is to explore the community service mechanism of numerical simulation and computing services for SMEs based on high-performance environment, reduce the threshold of using high-performance technology for SMEs, and establish a set of service specification for the needs of SMEs.

Abstract:In order to solve the problems of test and simulation verification in the development of UAV system, this study designs an integrative test and simulation system based on PowerPC hardware platform and VxWorks software platform. The system is based on the simulation computer. Considering the two functional requirements of integrative test and simulation, the hardware and software of the system are designed. This paper expounds the working principle and design idea of the system from four aspects: the overall development scheme, the hardware design and implementation, the system communication protocol, and the software design and implementation. Finally, different application environments are built to test and verify the system, and the data analysis proves the practicability of the system, which will effectively improve the efficiency of research and development of UAV flight control system.

Abstract:For traditional pipeline inspection systems that exist non-sufficient timeliness in the network data, spatial location relationship between networks is not intuitive and with poor continuity problems, this study uses the mixed reality, 3D modeling, and spatial positioning technology, in accordance with the need of pipeline inspection, dynamically constructs the pipe network and fuses with the actual features, intuitively shows the spatial location relationship between the networks and the real-time running condition of the pipeline network, designs and implements pipe network system based on mixed reality. Aiming at the problem of pipeline network model positioning, this study designs a spatial coordinate transformation method to complete the spatial mapping between virtual pipeline network and real pipeline network. For the construction and optimization of pipeline network model, this study proposes a dynamic construction and optimization method to realize the dynamic loading of pipeline network and the smoothing of local connections. The experimental results show that the proposed method can accurately integrate the virtual model with the real pipeline network, smooth the connection of the virtual pipe, and clearly and intuitively show the spatial relation of the pipeline.

Abstract:A large number of legacy serial codes need to be parallelized, and the complexity of parallel programs and the diversity of parallel computing platforms lead to high cost of transformation. For this reason, a three-layer parallel programming framework based on markup language is designed, which completes the two transformation stages from serial program layer to parallel intermediate code layer and from parallel intermediate code layer to target parallel programming language program layer. The method of language marking of serial code is used to realize the parallel intermediate code layer, which is actually an abstraction of the programming language of the parallel platform of shared storage and distributed storage. The framework also implements a performance tagging method, which can be used for automatic optimization of parallel parameters. The experimental results for radar data processing show that the corresponding parallel code is generated, and the parallel speedup ratio is equivalent to that of the manual parallel code.

Abstract:Aiming at the problems of low calculation efficiency and poor robustness of the existing face recognition system, this study proposed a face recognition system based on front and back interaction, including client, database, and server. First, a GrabCut-based facial Region Of Interest (ROI) extraction algorithm was proposed for the client end. Second, the extracted ROI is transmitted to the server, and the ResNet network is used on the server to extract facial feature points according to the ROI. Finally, the facial feature points extracted from the server were returned to the client, and the client performs Euclidean distance matching between this information and the feature points that were pre-stored in the database to obtain the face recognition result. The experiments were performed on the CeleA database and random videos, and the results show that the proposed ROI extraction algorithm significantly improves the accuracy and robustness of face recognition. Moreover, compared with the traditional non-interactive structure, the front and back interactive structure of the system greatly improves the computational efficiency of face recognition.

Abstract:The core system of the bank is an IT system mainly dealing with the most basic deposit and loan business of the bank, and the trading platform therein reflects the core of the bank and even the overall IT service capabilities. In order to improve the service capability based on the centralized architecture core system during peak trading hours, this study designs a core platform with transaction grouping service. The core platform covers a three-tier architecture of a trading server, a trading distributor, and a trading processor. The transaction grouping service algorithm is used to implement processing tasks grouped by transactions and dynamically allocate system resources. The experimental results show that compared with the traditional two-tier structure core trading platform of monitoring and distribution, the core platform with a three-tier structure with grouping services designed in this study has higher concurrent processing capabilities, and more stable system resource usage.

Abstract:In order to solve the increasingly prominent performance and safety problems of large-scale legacy systems, and to ensure the reliable and stable operation of the system, the performance and safety status of legacy systems are analyzed and studied, and an optimization strategy is proposed based on this. The optimization scheme is described from five aspects: front-end, database, server, system architecture, and system security. Finally, the optimization strategy is verified with examples. The verification results show that the scheme can effectively improve the performance and security level of the legacy system.

Abstract:Fusion experimental device is an important platform for fusion research. The performance parameters of the fusion device have important reference value for the fusion researchers to design the fusion reactor and select the experimental device. However, there is no integrated database for fusion experimental devices, which seriously affects the fusion research efficiency. In this study, a special database of fusion experimental devices was designed and implemented, which solves the problems of multi-source data regularization and heterogeneous data storage, and realizes the unified storage of relevant parameters. The functions of multi-field intelligent retrieval and multi-dimensional comparison analysis are developed, providing reliable and convenient data service for fusion researchers.

Abstract:The course of programming languages is usually the first core course for the majority of computer science. However, in the teaching practice, it is hard for students to master some complex and abstract knowledge in programming languages. To overcome the difficulties in teaching the course of programming languages, in this study, we design and implement a teaching system for the course of programming languages—a program dynamic analysis system. This system, by applying the knowledge of programming languages and some other core professional courses, implements program error detection and automated source instrumentation. We apply this system to the teaching practice, which helps students to understand and master the complex and abstract concepts in the course of programming languages, and their applications in real-world software development, thus to improve the teaching effectiveness.

Abstract:The traditional wireless fire warning system based on GSM costs too much and the battery on the sensing terminal drains so quickly that it has been difficult to obtain large-scale application. However, with the rapid development of LPWAN, the system cost and the energy consumption of the sensing terminal have reduced greatly, and the development of the intelligent fire warning system has been promoted quickly. In this study, a wireless fire intelligent early warning system is designed based on big data. Through this system, not only can the fire be discovered in time, but also the location of the fire can be grasped in real time. At the same time, the linkage between multiple departments and the individual can be realized. The actual results show that the system has low communication cost and long life of battery. Moreover, the system effectively reduces the incidence of the urban fire, as well as it has improved the efficiency of fire rescue and the level of fire service, which has been widely promoted and applied.

Abstract:Aiming at the problems that the current power grid information communication operation and maintenance system has hardware platform out-of-date service and information processing tasks cannot be completed quickly, this study proposes a migration model of micro service application system based on Spring Could to the cloud. This model migrates the power grid information communication operation and maintenance system to a cloud platform. In response to the management problems that occurred after the centralized system was migrated to the cloud and transformed into a distributed system, all system services were moved into the EDAS system to achieve one-click deployment, elastic scaling, grayscale release, and fault self-healing. In addition, a data consistency sampling verification method based on Gibbs sampling is designed to solve the problem of consistency check of huge amount of data in the cloud process. This method improves the efficiency of data consistency verification and reduces the cloud workload on the system. Afterwards, this article proves the system’s successful cloud deployment and reliable operation through the actual case of the cloud of the State Grid Anhui Electric Power SG-I6000 microservice system.

Abstract:In order to improve the accuracy of the human pose estimation task of convolutional neural networks, we propose an improved loss function based on Mean Squared Error (MSE) to deal with the pixel imbalance between foreground (Gaussian kernel) and background in heatmaps, assign different weights to the loss function according to different pixel values in the foreground and background, and named it Focus Mean Squared Error (FMSE). Compared with the mean squared loss function, the proposed focused mean squared loss function can effectively reduce the impact of pixel imbalance between foreground and background on network performance, help the network locate the spatial location of key points, improve network performance, and make the loss function converge faster in the training phase. Experiments are performed on public data sets to verify the effectiveness of the proposed focused mean square loss function.

Abstract:In order to find an optimal path with short distance and low node energy consumption in wireless sensor networks, an optimized ant colony algorithm DDEARA is proposed by using the “three-step progressive type” node finding method. Firstly, the dynamic radius search factor is used to find the next hop candidate nodes, which can ensure the convergence of ant colony algorithm and the uniform distribution of nodes location. Secondly, the node energy prediction factor is introduced to avoid the unreasonable phenomenon that the node is still overloaded when the energy is insufficient, that is, when all the energy of a node is consumed, all the data cannot be successfully transmitted. Finally, in the process of finding the next hop of candidate node, the direction factor is introduced, which has the directionality to find the node, avoiding the irrelevant node in the opposite direction to be selected as the next hop of candidate node, reducing the optimal path distance, saving node energy consumption, and improving the optimization efficiency of the algorithm. The simulation results show that DDEARA algorithm can realize the dynamic convergence of ant colony algorithm, the distance between adjacent nodes is moderate, the energy consumption of nodes is even, irrelevant nodes in the opposite direction are filtered, the optimal path distance is reduced, the optimization ability of algorithm is improved comprehensively, and the service performance and life of wireless sensor network are prolonged.

Abstract:Community detection based on the topological potential constructs the topological potential field by the link information of nodes, in which the community can be partitioned. However, there are a large number of isolated communities in the actual division process. The problem of community discovery with node attribute information, as an important part of the community, has become the main research direction of community discovery. This paper proposes a topological potential community discovery algorithm combined with label propagation (TPCDLP). First, combining the thought of label propagation, the attribute information is converted into the link weights between nodes. Second, the link weights are added to the topological potential to construct the topological potential field. Then, the subgroup communities are partitioned by the core node. Finally, the communities are partitioned by using the distance of the core nodes between the subgroup communities. Compared with six algorithms on three datasets with label attributes, the TPCDLP performs better on the improved modular degree $Q_{ov}^E$, information entropy $Entropy$, community overlap degree $Overlap$ and comprehensive index $F$.

Abstract:The traditional YOLOv3 network structure has poor robustness in extracting features such as over exposure or dark light, which leads to low recognition rate. A Dense-YOLOv3 model for traffic vehicle classification is proposed. The model integrates the characteristics of dense convolutional neural network DenseNet and YOLOv3 network, which strengthen the vehicle model feature propagation and reuse between convolution layers, and improve the anti-overfitting performance of the network. At the same time, the target vehicle is detected at different scales, and the cross-loss function is constructed to realize the multi-objective detection of the vehicle model. The model is trained and tested on BIT-Vehicle standard data sets. The experimental results show that the average accuracy of the model based on Dense-YOLOv3 vehicle detection reaches 96.57% and the recall rate is 93.30%, which indicates the effectiveness and practicability of the model for vehicle detection.

Abstract:The blockage of the high-speed railway segment may seriously affect the original timetable of the train. If it happens, the dispatcher needs to adjust the timetable. This study focuses on how to reschedule the timetable without changing the sequence of trains that are affected by the segment blockage. A train rescheduling model is formulated with aiming at minimizing the sum of the delay time of all trains in each station, and all constraints of the model is used to ensure the safety of trains. And a multi-stage variable neighborhood descent algorithm is proposed to solve the problem that it takes a long time to get the optimal or suboptimal solution. In the first and second stages of the algorithm, variable neighborhood descent combined with tabu table is used to quickly ascertain the trains which can maintain the original timetable after several adjustments, and then adjust other trains in the third stage. Finally, taking Xi’an-Chengdu Passenger Dedicated Line and the timetable of a single day as an example, the validity and real-time performance of the algorithm is verified by getting the timetable adjusted under various interval failure scenarios.

Abstract:Driver’s fatigue will affect the normal driving of the vehicle, and in serious cases will threaten the life safety of driver and passengers. Therefore, detecting whether the driver is fatigue can effectively protect people’s travel safety. In real scenario, generally, when the night light intensity is weak, the driver has a lot of time of fatigue driving, but the existing related detection algorithms cannot deal with the lighting problem, resulting in a low accuracy rate at night fatigue driving detection. Aiming at such problem, this study proposed a night-light fatigue driving detection algorithm based on low-light enhancement. Firstly, the LIME algorithm was used to perform low-light enhancement processing on the face image to improve the exposure of the image. Secondly, the face keypoint detection network was used to obtain the eye area of the image. Thirdly, the convolutional neural network was used to classify the eye area with open and closed eyes. Finally, the ratio of the number of eyes opened and closed per unit time is counted to determine whether the driver is in a fatigue state. The experimental results show that in the night environment, the detection algorithm proposed in this study improves the detection success rate by 15.38% compared with the existing algorithms, and achieves better results.

Abstract:Using the idea of quantum-inspired method, the normalized image pixels are represented in the form of quantum bits, taking full advantage of the strong correlation between a pixel and other pixels in its neighboring field, a spatial filtering image enhancement algorithm based on the quantum-inspired method is proposed, and the algorithm is simulated and compared for two spatial filtering templates: 3×3 and 5×5. Finally, image entropy is introduced to evaluate the image enhancement effect of the algorithm. In terms of the subjective vision and the objective evaluation index, the results show that the algorithm is superior to the traditional image enhancement algorithm. For the images with low visibility and contrast, compared with the size of 3×3 spatial filtering template, when the size of spatial filtering template is 5×5, the effect of the algorithm is better and the entropy of the corresponding result image is larger.

Abstract:In the age of the big data and artificial intelligence, Support Vector Machine (SVM) has been successfully applied in many aspects and becomes one of the common methods to solve classification problems. But the real world data is usually imbalanced, making its performance of classification significantly decreased. This study proposes to improve original standard Fuzzy Support Vector Machine (FSVM) by using inequality hyper-plane distance. The algorithm introduces parameter λ to controls the distance between hyper-plane and categories, and constructs fuzzy membership function by calculating sample mutually center distance, which can improve the falling precision of classification caused by imbalanced distribution of sample and noise data. The effectiveness of the proposed algorithm is verified by experiments, and the result shows that the proposed algorithm has a better effect of imbalanced data.

Abstract:How to measure the volume of express cartons simply and accurately is a difficult point in smart logistics. This study takes the express cartons in the logistics industry as the research object and proposes an algorithm for accurately measuring the volume of express cartons based on a 3D reconstruction network. To calculate the actual length of the carton, the image preprocess part segments the picture into express carton and background. Then the algorithm uses the 3D reconstruction network to obtain the voxel model of the express carton. Finally, removing the noise by the post-processing module, the algorithm gets the volume of the express carton. As the experiments show, the proposed algorithm improves the 3D reconstruction performance and obtains the volume of the express carton simply and accurately.

Abstract:Pedestrian detection is a current research hotspot, which has important applications in the fields of artificial intelligence system, vehicle assistant driving system, and intelligent monitoring. In the process of pedestrian detection based on HOG feature, the HOG feature is not obvious, the SVM classifier has high computational complexity, resulting in low recognition rate and high missed detection rate, this study proposes an improved enhanced HOG feature combined with the eXtreme Gradient Boosting (XGBoost) classifier for pedestrian detection. Firstly, the original image is preprocessed to get saliency map and HOG features. Then, the contrast of HOG features is enhanced and the pedestrian detection analysis is carried out with XGBoost classifier. Tested with the INRIA dataset, the experimental results show that the proposed algorithm has a significant improvement in recognition rate and detection speed.

Abstract:Chemical accident news data contains information such as news content, titles, and news sources. The text of news content is highly dependent on the context. In order to extract text features more accurately and improve the accuracy of chemical accident classification, this study proposes a Bidirectional LSTM (BLSTM-Attention) neural network model based on Attention mechanism to extract features of chemical news texts and realize text classification. The BLSTM-Attention neural network model can combine text context semantic information to extract text features of accident news through forward and reverse angles. Considering that different words have different contributions to the text in the accident news, the Attention mechanism is added to assign different weights to different words and sentences. Finally, the proposed classification method is compared with Naive-Bayes, CNN, RNN, BLSTM classification method on the same chemical accident news data set. Experimental results show that the BLSTM-Attention neural network model proposed in this study is better than other classification models in chemical data set.

Abstract:In order to improve the economy of distribution network operation and the reliability of power supply, the system average outage frequency and the system average outage duration are selected to represent the power supply reliability of the distribution network in this study, and the active power loss factor is considered at the same time, a multi-objective reconstruction model of distribution network is established, which takes the power supply reliability index into account. This study introduces quantum theory and Metropolis criterion into artificial swarm algorithm, and the optimal solution of multi-objective reconstruction model is determined by fuzzy satisfaction decision method, a multi-objective reconstruction model optimization method for distribution network based on improved artificial swarm algorithm is proposed. The distribution network reconstruction example simulation system established, and the feasibility and superiority of the reconstruction model and solution method are verified by comparison with other intelligent methods.

Abstract:Aiming at the problems of low accuracy, slow speed, and poor robustness of the current algorithm in solving the clustering problem, an improved butterfly optimization clustering algorithm was proposed. Based on the idea of elite strategy, the local search iterative formula of butterfly optimization algorithm was redefined, and then the selection, crossover, and mutation operations of genetic algorithm were fused. Test results on one artificial dataset and five UCI datasets demonstrate that the performance of the proposed algorithm is superior to other algorithms.

Abstract:The traditional Variational AutoEncoder (VAE) takes the flattened sample as input data directly. When the sample is image data, the effect of learning by this method is weakly. In this study, VAE with the convolution optimization is proposed to preprocess image data with multiple convolution networks of variable layers. Each convolution network sets different parameters to process the input data, then splices the results of different layers as the input of VAE. Clustering is implemented through the distance between the category label distribution of original dataset and the category distribution of each sample is calculated by adding a category encoder. The experimental results show that the convolution optimization method proposed in this study improves the clustering accuracy compared with the non-optimal VAE, increases the quality of the generated image and the diversity of the generated samples in the edge and shape.

Abstract:Medical records of patients are basic to the clinical diagnoses and treatments. Accurate recommendation of similar medical records can assist doctors in clinical decision making. In this study, we propose a new embedding model of medical records in real diagnosis and treatment scenarios. To recommend better medical records, we model the medical entities and their relationships in the medical records by heterogeneous graph embeddings. We conduct experiments on medical records of patients diagnosed with breast diseases from a Grade III-A hospital. The accuracy of the proposed model is improved by 2% compared with the existing model.

Abstract:In order to solve the Multi-Skilled resource-constrained Project Scheduling Problem (MSPSP), this study proposes an improved genetic algorithm. First, based on the mathematical model of the problem, a priority-based real number encoding method is established, and the objective function is converted into a fitness function for subsequent fitness calculations. Next, the niche technology based on group sharing is incorporated into the selection process of the genetic algorithm. In addition, with the help of deterministic sampling selection and subpopulation adjustment, the search ability of the algorithm is further improved. Then, gene repair and multiple verification mechanisms are introduced in the crossover and mutation operations to enhance the algorithm’s optimization ability. Finally, the overall process of the algorithm is given. The effect of the algorithm on the iMOPSE data set shows that the improved genetic algorithm is an effective method for solving MSPSP problem, and it has a sound reference significance for the study of related practical problems.

Abstract:Aiming at moving object detection, the current algorithm has certain applicability and limitations, as well as incomplete detection information. Based on the inter-frame difference method and hybrid Gaussian model, an improved hybrid Gaussian model target detection algorithm is proposed. To solve the problem of incomplete background contours of moving targets caused by the inter-frame difference method. This method is based on the traditional mixed Gaussian model, within a certain number of frames, checks the weights of all Gaussian distributions, deletes the Gaussian distributions that meet the conditions, and finally obtains a moving target with a clearer outline. The experimental results show that the algorithm in this study fully considers the influence of background characters on moving target monitoring. During the experiment, real grid data is used. The result shows that the proposed algorithm is 3.37% more accurate than other algorithms, with better accuracy and better adaptability to the environment.

Abstract:Fine-grained image classification is an important branch in the field of deep learning image classification. Since many different classified images are very similar in their features, and there is no particularly distinctive feature can be used to distinguish among them, it makes the classification task of fine-grained image more difficult than that of the general image. Therefore, a traditional image classification method needs to be optimized. Usually, visual and pixel-level features extraction is used in the training of the general image classification. However, direct application of this method to the fine-grained classification is not very suitable, and the effect still needs to be improved, while non-pixel-level features can be used to distinguish. Hence, we propose to combine text and visual information in the image classification, make full use of the features on the images, combine the text detection and recognition algorithms with general image classification methods, and apply it to the fine-grained image classification. In Con-text dataset, the experimental result shows that the accuracy obtained by the proposed algorithm has been significantly improved.

Abstract:The time-varying and space-varying characteristics of the marine sound field environment, the multi-source nature of the sound mechanism of underwater acoustic targets, and interference from other noise sources have brought many difficulties to the detection and identification of acoustic targets. Conventional target recognition methods are mainly based on the audio time-frequency domain analysis, it is difficult to obtain effective features and robust recognition effects. In order to solve these problems, transfer learning based acoustic target recognition is proposed. The pre-trained networks VGG and VGGish are used to extract deep acoustic feature analysis and fine-tune respectively. Experiments show that the proposed algorithm effectively improves the recognition accuracy and reduces the training time. The fine-tuned transfer learning algorithm has an average accuracy rate of 92.48% in acoustic target recognition, which achieved the state-of-the-art recognition result.

Abstract:At the evolution or maintenance stage of software engineering, software developers are required to handle many Software Change Orders (SCOs), which are generally prepared in natural language texts and involve one or more problem domains. The developers will accurately map these orders to corresponding source codes in the software programme and make ordered changes. This mapping requires creation of several search terms and search of them in the programme. Studies show that developers have difficulties in creating accurate and suitable search conditions for changes. In this study, the author proposes a TextRank-based search term recognition method for software change tasks. It enables identification and creation of search terms for software changes by analyzing tasks described in natural languages, to improve the accuracy, average precision, and recall rate of searching.

Abstract:Having studied the existing detection and classification algorithms, we design a scheme of fusion of improved Gaussian Mixture Model (GMM) and classification network (GoogLeNet) for vehicle detection and classification. In view of the inaccurate initialization and complex computation of GMM, we improve the algorithm of initialization models to increase the initialization efficiency. The five-frame difference method is used to execute the preliminary vehicle extraction. In the extracted vehicle area, GMM is used to get vehicle images, the five-frame difference method is combined with GMM to reduce the area of modeling and to increase the speed of vehicle detection and improve the real-time performance of the system. At last, we use GoogLeNet to execute the vehicle classification. The results show that the proposed methods have greatly improved the detection speed and recognition accuracy, and satisfy the real-time requirement of vehicle detection and recognition for surveillance video in real scenario.

Abstract:The traditional honeynet has many drawbacks such as inconvenient deployment, difficult flow control, and complex dynamic adjustment. This study proposes to use SDN, ODL, and Mininet technology to deploy lightweight virtual honeypots, build virtual honeynet topology, and use deep learning technology to optimize route selection. The experimental results show that the proposed SDN routing optimization mechanism has sound convergence and effectiveness, and can turn the attack to honeynet when the network is attacked, so as to reduce the damage caused by the attack and thus reduce the network attack threat.