Abstract:Deep learning has achieved remarkable results in target detection and classification when applied to computer vision. But in the field of object tracking, the target is only considered as a positive sample. Being lack of data support and more dependent on the location information, deep learning did not achieve remarkable effect in the object tracking field, while the traditional methods still occupy the main position. However, with the development of technology, deep learning has progressed greatly in the direction of object tracking in recent years. This paper introduces the basic concept and the main methods of target tracking technology. Combined with the development of deep learning in recent years in the field of target tracking, the emphasis is on the basic approach of target tracking technology with tracking by deep feature and tracking based on deep network and introduces the recently popular target tracking based on Siamese network in detail. At the end, the achievements of deep learning in the field of target tracking in recent years and future development of object tracking are summarized and prospected.

Abstract:In recent years, electronic data forensics has played an important role in the detection of cases. Because the electronic data have the characteristics of vulnerability, in order to finally analyze the useful evidence, it is necessary for the forensics personnel to have professional electronic forensics technology and methods to ensure the authenticity and objectivity of the cases. Three kinds of forensic techniques are analyzed in detail:electronic forensics based on Windows, electronic forensics based on smart phones, and electronic forensics based on network. Electronic forensics based on smart phones, include Android mobile phone and iPhone mobile phone. And the future development direction of electronic forensics technology is put forward as well.

Abstract:In recent years, fuzz testing has become one of the most popular and efficient methods to detect program bugs and vulnerabilities. By mutating seed files, fuzzing tools generate a large volume of test inputs, and feed them to the program under test in order to expose security weakness. Current researches mostly focus on improving the mutation algorithm to make newly generated files cover more target program codes. However, little attention had been paid to elaborately optimizing the policies to sort seed files to be fuzzed, which prioritizes the seed files with higher probability to cover new program spaces in the fuzzing process, and consequently improves the efficiency of fuzzing. Therefore, we proposed a coverage frequency based selection approach to guide the fuzzer to execute promising seed files first. To do so, we first kept tracking how many times each edge between two basic blocks has been executed by the target program in the fuzz testing. Based on the number that each edge has been executed, we then categorized them into high-frequency edges and low-frequency edges. Only seeds containing more low-frequency edges, as well as being executed very fast by the target program, were assigned with high priority. We implemented our method on American Fuzzy Lop (AFL), one of the most popular fuzzers and applied the modified AFL version to 5 real world programs. The result shows that our approach can improve both the efficiency of AFL and the code coverage explored in the target program.

Abstract:In order to effectively evaluate DDoS attack and defense behavior to defend against DDoS attacks, this study first analyzed the current research status of DDoS attack and defense evaluation, and then established DDoS attack and defense behavior confrontation net based on stochastic Petri net. The attack and defense steady state probability is used as the basis for the evaluation of attack and defense behavior. The solution of attack and defense game strategy based on the attack and defense game theory were proposed. In the end, we carried out the stability analysis of the DDoS attack and defense behavior confrontation net, and comprehensively considered the factors of attack and defense behavior gain and attack and defense behavior intensity to simulate and evaluate. The evaluation results show that the method is more reasonable and pertinent.

Abstract:In view of how to quickly find Chinese printed text image files in bulk image file for Optical Character Recognition (OCR) recognition in practical application scenarios, this study designs heuristic rules for the inherent characteristics of Chinese text, based on the Stroke Width Transform algorithm (SWT), and combines horizontal projection technology with discrete Fourier transform, a Chinese printed text image file recognition technique suitable for tilt angles between -90 and 90° is proposed. The experimental results show that in 1606 test set image files, the overall recognition F-measure of the algorithm for text image files is 0.95, and the average recognition time is 0.65 s.

Abstract:Binary networks have obvious advantages in terms of speed, energy consumption, and memory consumption, but they cause a great loss of accuracy for the deep network model. In order to solve the problems above, this study proposes a staged residual binarization optimization algorithm for binary networks to obtain a better binary neural network model. In this study, we combine the random quantification method with XNOR-net, and propose two improved algorithms, namely applying weights approximation factor and deterministic quantization networks, and a new staged residual binarization BNN training optimization algorithm, in order to obtain the recognition accuracy of the full-accuracy neural network. Experimental results show that staged residual binarization algorithm can effectively improve the training accuracy of binary model, and does not increase the computational complexity of the related network in the testing process, thus maintaining the advantages of high speed, low memory usage, and small energy consumption.

Abstract:Based on the filter of meteorological and hydrological stations and according to the cloud distributed storage technology and meteorological data processing technology, the main idea, function structure, and product application of the system design are introduced. The real-time collection process of precipitation data is combed, and the data processing and sharing system are established. The system integrates the rainfall data of the Yangtze River from meteorological and hydrological departments, and realizes the efficient data exchange and sharing application with hydrology department.

Abstract:IPoIB is a protocol that supports traditional Ethernet over InfiniBand networks, allowing IP applications to run on InfiniBand networks. We realize IPoIB on the domestic parallel system to improve the performance of IPoIB on the system, we also propose four optimization methods, namely, reordering packets, optimizing memory copy, tuning network parameters, and avoiding delayed acknowledgement. Practice shows IPoIB runs correctly on the domestic parallel system, and under the methods of optimum, IPoIB's network bandwidth performance is nearly 6x higher than the not-tuned version. IPoIB has more advantages compared with 10 GbE, and the reordering method shows obvious effects.

Abstract:Chinese insurance market is experiencing the rapid development period, the variety of customer, insurance stakeholder, and product makes the product recommendation become a hot issue. However, accurate product recommendation faces the technical challenge brought by privacy-preserving and trustworthy problem. Firstly, this study analyzes the matching problem of customer demand and insurance company. Then based on blockchain technology, a new insurance product recommendation model is proposed, customer and insurance company can submit the privacy information needed by the other party to the recommendation model for requirement matching, thus more accurate product recommendation result can be achieved. The experiment shows that the model can achieve the security and transparency of the product recommendation process while preserving the privacy of participants.

Abstract:With the continuous development of Internet technology, and the popularization of mobile phones, computer tablets, and other mobile terminals, live video streaming has flourished and expanded over the past few years. Almost every live video streaming platform in China has virtual-gifts donating mechanism, which allows viewers to buy virtual-gifts provided by the platform for rewarding the broadcasters. Viewers' virtual-gifts donation is one of the most important sources of revenue for both the broadcaster and the platform, which makes it important to understand the viewer's behavior, so that it can be used to explore user's value and enhance user's liquidity. In this study, we take Douyu live video streaming platform as a case study, mainly focusing on the high consumption community on the platform. We specifically construct viewer features to analyze their behavior through clustering approach. The experiment result shows that the high consumption community can be divided into three clusters with significant difference in their behavior. We also conduct detailed analysis regarding viewer characteristics for all these three clusters, and offer suggestions for the platform to provide diversified user-oriented services.

Abstract:The voter in the mimic Web server calculates the similarity of the heterogeneous executor response webpage in order to judge whether the response is legal output and thus to prevent webpages tempering. At present, the voter treats entire Webpage as a string and uses the string edit distance to calculate the similarity of the webpages. In this way, it caused problems such as large amount of calculation, ignorance of the original structure information of the webpages, and so on. In this study, the improved simple tree matching method is used to calculate the similarity of the webpages by calculating the similarity of the DOM tree of the webpages. The matching degree of DOM tree node is determined by the editing distance of the node string. The proposed algorithm is applied to the mimic Web server to verify the webpage tamper. Compared with the existing algorithms, the algorithm used in this study not only adapts itself to the heterogeneous but also improves the efficiency and accuracy of the voter.

Abstract:In order to break the traditional way to update the LED content and data transmission range restrictions, the Tencent QQ link platform and Light Emitting Diode (LED) dot matrix display are combined in this study, an innovative LED display is designed. In this design, users can send the text message by QQ client in the mobile phone, and the message is sent through the software of QQ client, QQ cloud, and the QQ board, and then displayed in the LED display after series of message processes and communications in LED board. Therefore, the information in the LED display can be updated in real time. This design is the typical application of the Internet of Things (IoT) technology, and is valuable in terms of intelligent hardware design.

Abstract:In order to solve the problem in bridge management of real-time security monitoring and expected safety prediction based on the realization of TMS320C6748 laser spot center positioning, a bridge deflection monitoring and prediction system based on Android is designed. The monitoring system adopts a modular design idea and consists of a Socket server module, a Web server module, and an Android module. Combined with the real-time bridge deflection data received from the Socket server module, the database connection and operation are established, and the information environment combined with the mobile terminal and the Web server is used to realize the function of bridge information annotation, search, and deflection information display based on Baidu map API. The prediction system predicts the short-term bridge state changes by comparing and analyzing the comprehensive scoring method and the optimized Markov chain. After many times of simulation test of the simulated bridge, the prediction result of the algorithm is consistent with the actual situation, and the running time of the whole software is controlled in 4 to 6 seconds, realizing the requirement of real-time performance. The results show that the system has strong accuracy, real-time, and practicability, which provides a convenient and accurate way for the bridge managers to supervise the bridge and timely formulate the bridge concrete maintenance plan.

Abstract:Object detection plays a key role in robotics, and as one of the most popular platforms for robot development, it is very necessary for ROS (Robot Operating System) platform to achieve fast and accurate object detection. Recently, the deep learning method is the core technology to realize the object detection function, but the object detection packets carried by the ROS platform are still based on the traditional local image feature description method, which with poor robustness and weak generalization ability. Aiming at the above problems, we propose a customized object detection model based on SSD_MobileNet framework, which combines the image dataset independently, and integrate the model to the ROS platform to achieve fast and accurate object detection function.

Abstract:In practical field geological survey, the use of recording paper to record measurement data has certain limitation. It not only has a heavy workload and is error-prone, but also is inconvenient to carry measure instruments. Therefore, it restricts the measure efficiency for civil engineers greatly. Considered the advanced technology and portability advantages integrated with smart phones, this study utilizes the Java language to design a geological-measure electronic recorder App system based on Android platform. In the concrete development process, the SQLite database is first used to save the relevant data obtained by the project. Then, the Baidu map is integrated to collect the work point positioning information. Furthermore, the gyroscope is employed to measure the dip and dip angle. Finally, the socket technology is used to connect the hardware equipment and collect the experimental data of rock charge load test. The electronic recorder can realize the automatic recording and correction of the multimode measurement data, and improve the field measure efficiency for civil engineers.

Abstract:Although deep learning based vehicle detection approaches have achieved remarkable success recently, they are still likely to miss comparatively small-sized vehicle. To address this problem, we propose a novel Dilated Fully Convolutional Network with Grouped Proposals (DFCN-GP) for vehicle detection. Specifically, we invented a grouped network structure to combine feature maps from both lower and higher level convolutional layers for the generation of object proposal and focusing more on lower level features, which are more sensitive to discovering small object. In addition, we increase the size and reception field of the feature map in the last convolutional layers to keep more detailed information via dilated convolution, which is used in both object proposal and vehicle detection sub-networks. In the experiment, we conducted ablation studies to demonstrate the effectiveness of the grouped proposals and dilated convolutional layer. We also show that the proposed approach outperforms other state-of-the-art methods on the UA-DETRAC vehicle detection.

Abstract:In order to reduce the probability of accidents caused by bad lane changing behavior, it is necessary to identify the lane changing behavior during the actual driving of the vehicle. In this paper, the IOS intelligent device is used to collect data, and the corresponding feature vector is established. The vehicle lane change behavior recognition model based on support vector machine is proposed. An improved N-δ sliding window interception algorithm is proposed for the recognition of continuous lane change behavior so as to divide the data containing multiple behaviors quickly, the sample data is used to verify the feasibility of the N-δ sliding window interception algorithm and the validity of the classifier.

Abstract:For flexible shop scheduling in the case of machine resources and processing route selectable, a flexible shop model is established with minimum and maximum completion time and time penalty costs as targets. According to the characteristics of the problem, an improved firefly algorithm is proposed. The algorithm designs a coding strategy with greedy ideas. A firefly individual represents the processing sequence and process processing position. It adopts an adaptive selection strategy to adapt the step length and improve the accuracy of the algorithm. The introduction of POX cross strategy to improve the algorithm's local and global optimization capabilities, and the use of greed to improve the convergence speed of the algorithm. The performance of the algorithm is verified by comparison with example simulations and algorithms. The experimental results show that the improved firefly algorithm is effective for solving the flexible shop scheduling problem.

Abstract:Person re-identification aims at pedestrian target matching under distributed monitoring systems. Compact and robust feature is critical to it. For this reason, this study proposes a feature extraction method based on feature fusion network. Firstly, the STEL algorithm is used to enhance the immunity of LOMO feature to background noise, and the KPCA algorithm is used to reduce dimension. Subsequently, we explore the complementarity between manual features and Convolutional Neural Network (CNN) features, and integrate the improved LOMO feature into the CNN to obtain a fusion feature with better performance. Experiments on two datasets (VIPeR and CUHK01) validate the effectiveness of our proposal, the Rank-1 of fusion feature is 3.73% and 2.36% higher than the cascade feature, respectively.

Abstract:Aiming at the problem of large deviation in existing PM2.5 concentration prediction, a novel model based on Improved Firefly Algorithm optimization SVM (IFA-SVM) was proposed. In this model, two neighborhood search strategies and variable step size mechanism were employed to improve FA. The IFA was applied to optimize the SVM parameters (C,, and), and an outstanding model was constructed to forecast PM2.5 concentrations in Taiyuan. The neighborhood search strategies can provide better candidate solutions; search step size was dynamically tuned by using variable step size strategy to accelerate convergence and obtain a trade-off between exploration and exploitation. The performance of the proposed IFA-SVM model has been compared with FA-SVM, Genetic Algorithm (GA)-SVM, and Particle Swarm Optimization (PSO)-SVM. Experimental results show that the proposed IFA-SVM model has achieved more accurate performance for PM2.5 forecasts in 1 day ahead and 3 days ahead compared to other method.

Abstract:In view of the fact that the building area under construction has features that are different from the features of the surrounding non-construction buildings and different from the surrounding natural environment texture features, a construction area recognition method based on the color and texture features of the building under construction is proposed. Firstly, color and texture features are extracted from images that only contain the image data of the building under construction. The image feature index database is constructed from these feature vectors. Then, the image to be detected is divided into blocks, and the color vegetation is masked to the green vegetation area and the feature vector is calculated. It is measured by the similarity with the feature index database, to determine the position of the image block in the entire detected image, and to select the red rectangle and the unique identifier box for the detected building under construction. The experimental results show that the proposed method for recognizing the building area under construction can effectively identify the building area under construction in urban images from high altitude. The system based on this algorithm can be applied to urban planning.

Abstract:With the development of deep learning, many complex problems in semantic segmentation tasks are solved, which lays a solid foundation for image understanding. The proposed algorithm highlights two aspects. Firstly, our algorithm fuses multi-scale features from different levels of deep convolutional network by using multi-level deconvolution network. Then our algorithm upsamples these feature maps by deconvolution, meanwhile zooms them up to the original image size to predict semantic categories pixel-to-pixel. The second one, we propose a new method for data processing which is batch centralization algorithm, in order to improve the performance of network structure in this study. Through experimental verification, the mean IoU of semantic segmentation on the SIFT-Flow dataset reaches 45.2%, and the accuracy of geometric segmentation reaches 96.8%. The mean IoU of semantic segmentation on the PASCAL VOC2012 dataset reaches 73.5%.

Abstract:This study proposed a building energy consumption prediction method based on Q-learning algorithm. By modeling the building energy consumption prediction problem as a standard Markov decision process, combining with the deep belief network to model the state, we use Q-learning algorithm to achieve the real-time prediction of the building energy consumption. Based on the building energy consumption data published by Baltimore Gas and Electric Power Company of the United States, the proposed model were tested and the results show that the Q-learning algorithm can be used to predict the building energy consumption successfully. Moreover, deep belief network can improve the prediction accuracy effectively. In addition, some experimental results further verify the influence of related parameters on experimental performance.

Abstract:Text sentiment classification is a hot topic in the field of natural language processing. One of its important applications is to dig out important information from online comments and grasp the trend of public opinion on the Internet. Therefore, this study proposes a method of text sentiment classification based on GDBN neural network. The algorithm improves the hidden layer in the DBN neural network by introducing genetic algorithm, which is of powerful global searching ability, and the algorithm optimizes the number of hidden units and obtains the appropriate value of the current model, then the modeling and feature extraction of this model. Finally, we can classify the extracted features of the BP neural network. By testing multiple data, the results show that the proposed algorithm is effective.

Abstract:For the problem that when the user score is not enough, the recommender system significantly reduces the data sparse recommendation performance, a Collaborative In Deep Learning algorithm (CIDL) is proposed. The algorithm firstly conducts the deep learning on a large amount of data, and then performs collaborative filtering on the rating (feedback) matrix to arrive at a recommendation item for the user. This study uses real movie data to test and to compare it with the other four excellent algorithms. It is proved that CIDL can effectively solve the problem of reduced performance due to data sparseness and improve the accuracy of the recommendation.

Abstract:After crawling to obtain a data set of large web pages on a large commodity site, the data set is screened to further get the target data set. Before screening, preparation must to be done is to delete the redundant tags in the web pages. Therefore, the algorithm of deletion tag is given with the idea of a recursive algorithm. The design idea of tag deletion function is put forward. 2 time design improvements are carried out to optimize the performance. Finally, the design idea of dual thread is adopted. The dual threads are 1 maintain buffer thread and 1 tag deletion thread. In single computer environment, experiments show that the optimized tag deletion function only takes 19.7 seconds for each 1000 pages, and only 1.1 hours for 200 000 web pages.

Abstract:Aiming at the problem that the overload of information resources of TV products leads to the difficulty of user selection, this study mainly focuses on the improvement and application of article-based collaborative filtering algorithm in television product recommendation system, and combines the personalized recommendation technology with TV product system to meet the need of users and operations. In the recommendation process, the user's preference data model is first collected, and the duration of the user watching the television product is taken as the explicit characteristics of the user's preference. Then, it is improved by introducing the weight of on-demand amount in the traditional collaborative filtering algorithm, and the Euclidean distance method is used to calculate the similarity of the items. Finally, the viewing time of the target user on the television products is predicted according to the neighbor set, and a recommendation result is obtained. Experiments show that the introduction of on-demand amount weights can improve the accuracy of recommendations.

Abstract:Software testing plays a vital role in producing reliable software. The debugging of software is divided into two steps:fault localization and fault modification, where the fault localization is the most time-consuming and tedious work. Generally, most of test cases in a test suite performed successfully. In order to improve the availability of the defect location method based on the program spectrum, the method should have the ability to adjust the weight of successful coverage automatically. That is, if the contribution of the number of statements to the suspiciousness is reduced gradually with the increase of the number of successful test cases, the effectiveness of fault localization method can be improved greatly. Based on this idea, this study proposes an EPStar (EP*) defect location method, which can effectively adjust the effect of successful use cases, so as to avoid the excessive influence of the number of successful use cases to the defect location effect. To improve the accuracy of error location, the experimental comparison shows that EP* method has higher defect location accuracy than several existing defect location methods.

Abstract:Word vector representation is a sound way to catch the grammatical and semantic information of words. In order to improve the accuracy of the semantic information of the word, this study proposes an improved training method model based on the GloVe by analyzing the characteristics of the co-occurrence matrix and using the distributed hypothesis. This method summarizes the general rules of irrelevant words and noise words in the co-occurrence matrix from analyzing the word frequency of Wikipedia statistics. Finally, we give the evaluation results of word vector in word analogy dataset and word correlation dataset. Experiments show that the method presented in this paper can effectively shorten the training time and the accuracy of the word semantic analogy experiment is improved in the same experimental environment.

Abstract:Aiming at the high sparsity problem of user's check-in data and user privacy in LBSN, a hybrid recommendation model (SoGeoCat) is proposed. Firstly, the user's potential point-of-interest is learnt from the user potential point of interest data model. Secondly, the user's potential point-of-interest is incorporated into a category based matrix factorization model and then optimized. Finally, the proposed recommended strategy is according to the user and feature matrix and the point-of-interest matrix. Based on the Foursquare real dataset, the experimental results show that:(1) compared with several other recommended models, the algorithm fills the user's potential point-of-interest into the matrix, which can effectively alleviate the impact of data sparsity; (2) the algorithm can protect the user's family information; (3) the influence of the category information in the recommendation model can improve the recommendation effect.

Abstract:In order to solve the problem of manually annotating large-scale corpus, this study, based on the model of Probabilistic Latent Semantic Analysis (PLSA), proposed a method of automatic emotional annotation for news comments. First of all, the "doc-topic" and "word-topic" probability matrixes were computed by PLSA model. Then, drawing upon the "word-topic" together with the ontology lexicon, the emotional categories of the topics were annotated, with the presupposition that the emotional category of words is similar to those of words within the topic which occurs most frequently. Finally, the automatic annotation was made via the "doc-topic", with the assumption that the emotional category of topics is equivalent to those of topics within the document which occurs most frequently. The experimental results showed that the accurate rate of the method proposed by this study reached about 90%.

Abstract:In order to realize automatic detection of bearing dust cover in industrial field, the key technology of machine vision detection is studied, and a set of bearing dust cover detection is put forward based on machine vision. According to the characteristics of the bearing image, based on image filtering, a least square method is used to carry out circle fitting circle detection algorithm, the bearing detection and extraction of the ROI area are completed. The characteristics of the connected region are analyzed, then binary and morphological operations are carried out, and finally quantitatively processed and judged. The experimental results show that the method has good real-time performance and high detection accuracy, and can meet the industrial requirement of automatic detection of the bearing dust cover.

Abstract:With the development of society and the advancement of civilization, human beings are increasingly demanding the level of intelligence, convenience, and safety of vehicles. In view of the problem of less parking lot and difficult parking, the waiting time factor is added to the A* algorithm, and the waiting time is added to the number of heuristic functions, the distance of the comprehensive path and the waiting time are two factors to plan the best path for the tasks of parking into the lot and driving out of the lot, and a three-dimensional A* intelligent parking algorithm is designed. Under the structured environment, the optimal parking space is reserved according to the given parking lot map, the optimal path of the Automated Guided Vehicle (AGV) is planned for multiple simultaneous tasks, and the AGV vehicle is arranged to go to the designated parking space or to move out of the lot, with minimized waiting time of the customers to park or take the car, and making the total cost minimum, presenting a truly "Internet-enabled" smart parking experience.

Abstract:In this study, we use the method of textual emotion analysis to study the user satisfaction. First of all, we use the LDA model to establish the structure model of the user satisfaction. Then, the emotion tag of sentence is extracted based on dependency parsing technology, and the HowNet sentiment dictionary is combined with the semantic similarity algorithm to identify the sentiment tendency of the sentence. Finally, the fuzzy comprehensive evaluation method is used to study the user satisfaction. Taking Mobike for example, the research shows that the user satisfaction of Mobike bicycle is higher as a whole. However, the following phenomenon affects the improvement of customer satisfaction of Mobike, namely, required deposit is high, deposit refund is not timely, fault car is more, and the accuracy of positioning software is low.

Abstract:In the past decade, the incidence of heart disease has been on the rise and remains high in the world. If the physical examination indicators related to the human body can be extracted by computer measures, and the influence of different characteristics and their weights on heart disease can be analyzed through machine learning, it will play a key role in predicting and preventing heart disease. Therefore, a prediction method based on clustering and XGboost algorithm is proposed in this study. By preprocessing the data and distinguishing the features, the data sets are clustered by clustering algorithm, such as K-means. Finally, the XGboost algorithm is used to predict and analyze. The experimental results show that the proposed method based on clustering and XGboost algorithm is feasible and effective, which provides accurate and effective help for the application of medical recommendation.

Abstract:To solve the problem of performance testing for gap of network security special products, a test method for gap performance was studied. Two test methods including the use of iPerf software and IXIA hardware were designed. Both test methods can be applied to the test of gap performance. At the same time, a test method for the performance requirements of gap in the national standards was also analyzed, and a comparative analysis was made with the performance requirements of network security special products. Finally, the throughput and system delay performance for the gap of network security special products are tested through experiments. The distribution results of the 14 gap products are given.

Abstract:Image semantic segmentation is one of the most important research directions of computer vision. Compared with traditional algorithms, image segmentation based on deep-learning performs better, and can be applied to the scene understanding stage of traffic monitoring and automatic drive. However, the speed of complex segmentation network on embedded platform is too low to be practically applied. Therefore, in view of the application of traffic monitoring and automatic drive, the image segmentation network based on deep convolutional encoder-decoder architecture was used to complete the road scene segmentation on the embedded platform NVIDIA Jetson TX2. Meanwhile, in order to accelerate the network, the model was simplified and transformed to engine based on TensorRT2 provided by NVIDIA, which including plugin layers adding and CUDA parallel optimization. The experimental results show that the speed-up ratio can reach ten, which provides support for the application of the complex structure segmentation network on the embedded platform.

Abstract:The recognition of the named entity in the judgment documents is the key step in the automatic trial. How to effectively distinguish the key named entity of the case is the key point in this study. Therefore, this study proposes a neural network model based on SVM-BiLSTM-CRF for property dispute of trial cases. First, the sentences containing the key named entities are selected by SVM, and then the sentences are converted into the character level vectors as input, and the BiLSTM-CRF deep neural network model suitable for the identification of the property dispute referee's named entity is constructed. By constructing training data for verification and comparison, the model shows higher recall and accuracy than other related models.

Abstract:Timely diagnosis and intervention for potential diabetic retinopathy patients is very positive in improving the overall visual quality of diabetic patients and reducing medical costs. Because the fundus fluorescence images of preclinical diabetic retinopathy and normal people have no obvious difference in visual perception, this study recognizes the two groups of images through the widely used texture feature algorithm and support vector machine. Through the 10-fold cross validation of 185 fundus autofluorescence images, the LBP algorithm has a sound recognition effect. The 10-fold cross-validation accuracy of the 59-dimensional LBP operator with "Uniform" patterns reaches 91.89%. And in the case that the test set and the training set are randomly divided by 1:1, the recognition accuracy of 92 fundus fluorescence images in the test set reaches 88.12%, and the AUC is 0.943.

Abstract:The Android PowerManager provides a programming interface called WakeLock to protect critical computation from system hibernation. The WakeLock acquisition and release in the application can have a significant impact on power dissipation. However, the misuse of WakeLock increases the energy dissipation of the device and seriously affects the user experience. This article analyzes and summarizes the types and causes of WakeLock misuse. Furthermore, we developed an Android application that can detect and optimize the WakeLock misuses to reduce the power dissipation.

Abstract:Bayesian network performs probabilistic inference for network model by determining variable node network structure and parameter learning, under the condition of sample data is not too big, an accurate prediction results can be obtained. The training sample data are selected from each data platform for the standardization of college and university student behavior, which is used to build a Bayesian network and to learn the parameters by the network to get the inference model, and then the poverty status of college students is predicted by the model. The predict results show that there are no significant differences between the predict results and the actual samples. Thus the poverty level of college student can be accurately determined by data analysis.

Abstract:It was found that the process of carrying, storing, and forwarding data of sensing nodes ignored the content filtering of information carried by nodes by analyzing the cooperative process of mobile crowd sensing. As for purposeful data acquisition, this method of collecting data first and screening data later spent more time in analyzing and screening data. At the same time, the proportion of efficient data was not high. Taking these into account, a node selection mechanism based on sentiment text data screening in mobile crowd sensing environment combining genetic algorithm is designed. In the node selection mechanism, the perceptual nodes are selected by screening the data types so as to obtain the emotional text data of the mobile users in the perceptual environment. The experimental results show that the efficiency of data processing is increased by 27.6%, and the proportion of effective data is increased by 21% by using this method. Therefore, the method proposed in this study can effectively improve the efficiency of the whole data processing.

Abstract:NLTK is a module for processing natural language text in Python, but it has limitations when processing Chinese text. To extracted information in the text by using NLTK, the means created in this study included a group of methods, such as common context words extraction, bigrams words extraction, probability statistics, and discourse analysis. Both of NLTK text content extraction framework suitable for Chinese texts and implementation method are obtained. In the results of empirical, it finds the content of the corpus which reflects the characteristics of the text, and gets the conclusion that a strong correlation between the results of extraction and text topic.

Abstract:Deep learning is a kind of machine learning method widely used in the field of artificial intelligence. The high dependence of deep learning on data makes the dimension of the data needed to be processed, which greatly affects the computing efficiency and the performance of data classification. Taking data dimensionality as the research goal, the methods of dimensionality reduction in deep learning are analyzed in this paper. Then, taking Caltech 101 image dataset as experimental object, VGG-16 depth convolution neural network is used to extract image features, and PCA statistical method is taken as an example to achieve dimensionality reduction of high-dimensional image feature data. Euclidean distance is used as a similarity measure to test the accuracy index after dimensionality reduction at the testing stage. The experiments show that image can still maintain high feature information using the PCA method to reduce the data dimension to 64 dimensions after extracting the 4096 dimensional feature of the fc3 layer of the VGG-16 neural network.