Abstract:Since the proposal of generative adversarial networks, GAN has become a research hotspot of artificial intelligence. GAN adopts the method of the zero-sum game between two people, which consists of a generator and a discriminator. The generator is responsible for generating the sample distribution, and the discriminator is responsible for determining whether the input is a real sample or a generated sample. The generator and discriminator constantly interact and optimize to achieve the optimal effect. The model of GAN is undoubtedly very novel, but there are also many shortcomings, such as the problem of gradient disappearance, collapse mode, and so on. With the deepening of research, GAN has been continuously optimized and expanded, and the derivative models of GAN have emerged in endlessly. GAN has been optimized and improved. Also, GAN can be applied in different fields, though it is mainly used in the field of computer image and vision. It has outstanding effects in the field of image. It can generate high-resolution realistic images, repair images, transfer styles, and generate video and prediction. GAN can also generate text, to do some work, such as dialogue generation, machine translation, voice generation, and so on. GAN is also involved in other fields, such as generating music and decoding codes. However, the application effect of GAN in other fields is not significant. Therefore, how to improve its application effect is worthy of further study, which will make the generation of confrontation networks shine in artificial intelligence.

Abstract:The narrow-band method is a common acceleration method for level set image segmentation. The traditional narrow-band still has redundant computational regions; When the traditional narrow-band method is combined with the LATE (Local Approximation of Taylor Expansion) level set model, the image segmentation efficiency may be reduced. In order to solve these problems, a rectangular narrow-band method based on LATE level set image segmentation model is proposed in this study. The level set is subjected to the following narrow-band processing before each LATE level set iteration. First, find out all the points of zero crossings of the level set; second constrict the points of zero crossings by the activity constraints, eliminate the inactive points of zero crossings, and effectively reduce the area of the narrow-band, then generate a rectangular narrow-band for the points of zero crossings by the active constraints, optimize the overlapping rectangular narrow-band so that the total area of the rectangular narrow-band is as small as possible. Finally, the level set differential equation is solved in the narrow-band of the rectangle, and the level set is updated to complete this iteration. In the different stages of the level set evolution, the area of the traditional narrow-band and the rectangular narrow-band of this study are compared. As the number of iterations increases, the ratio of the area of rectangular narrow-band to the area of traditional narrow-band is gradually reduced to zero, indicating that the rectangular narrow-band method effectively reduces the amount of redundancy calculation. For images with different degrees of intensity inhomogeneity, the proposed method is compared with the LATE method, the direct narrow-band method, and the DTM narrow-band method. The direct narrow-band method and the DTM narrow-band method have lower segmentation efficiency than the LATE method, and the segmentation quality is greatly affected for some images with severe intensity inhomogeneity. Under the condition of maintaining good segmentation effect, the segmentation speed of the proposed method is faster than that of LATE method. The rectangular narrow-band method in this study effectively reduces the complexity of the algorithm and improves the efficiency of image segmentation.

Abstract:With the significant increase in the number and frequency of private cars in cities, the problem of "parking difficulties" has gradually become a "bottleneck" that restricts urban development. In order to make reasonable use of the city's limited parking resources, the best way is to establish a city-level parking guidance system. At this stage, there is no effective solution. The reason is that the cost of obtaining parking data is too high. Therefore, how to reduce the procurement cost without affecting the accuracy of the parking data becomes the key to solving the problem of "parking difficulties". First, based on the spatiotemporal sensitivity of parking data, parking lots with significant data differences are divided into different clusters. After verifying that parking data in the same cluster complies with Pareto's principle, the top 20% of the most influential parking lots are selected as sample parking. At the site, sensors are installed to obtain real-time parking data and used as sample data. Considering that the patch data obtained by the existing algorithm is not satisfactory, this study upgrades the one-dimensional parking data to two-dimensional, and uses the improved Deep Convolution Generative Adversarial Networks (DCGAN) to generate new data settings and sample data. Roughly the same, any new data set can be used as any missing parking data in the same cluster. The implementation results show that the proposed scheme in this study can not only obtain a large number of high-quality "pseudo-data" in batches under the condition of limited perception, greatly reduce the acquisition cost of parking data greatly, but also significantly improve the repair effect compared with the current research.

Abstract:The mobility scenes of urban transport applications in the vehicular Ad Hoc networks is analyzed in this study. Based on the VanetMobiSim traffic simulator and NS-2 network simulator, two kinds of typical scenarios of urban intelligent traffic were built, which are Models of junction and two-way fast four-lane, and the AODV and DSDV routing protocols were simulated and analyzed in these two scenarios simultaneously. Four evaluation indicators which include end-to-end delay, jitter rate, packet loss rate, and control packet overhead, are introduced to evaluate and compare the performances of different protocols in the same scene. In addition, in the simulation experiments of AODV, the effects resulted from vehicle speed, vehicle density, the maximum number of connections, and the number of packets sending per unit time are also analyzed. The simulation result shows that selecting the appropriate mobility scene model helps to evaluate the performance of the protocols objectively, different protocols present different performance and adaptability in different scenarios. DSDV performs better when applied to simple and stable mobile scenarios, while AODV has better adaptability to complex and ever-changing mobile scenarios. Moreover, AODV protocol is greatly affected by the complexity of network topology and the frequency of the change in structure.

Abstract:The factors affecting the accuracy of real-time data of HPIC dose rate in nuclear radiation monitoring stations are complex, such as natural factors of rainfall, temperature and humidity, wind direction and solar radiation, objective factors of equipment anomalies and radioactivity, etc. When it is found that the radiation monitoring state is abnormal, it is difficult to analyze the cause of the deviation of the monitoring data. Combined with the monitoring data of massive historical radiation series of ERMS, the characteristics of rainfall, temperature and humidity, air pressure, wind direction, electrons in the zenith direction of solar radiation and the radiation values of surrounding sites are deeply explored. HPIC is established based on the Gradient Boosting algorithm (referred to as GB algorithm). The online prediction model of dose rate radiation data effectively combines the natural characteristic factors, reduces the natural factor's analysis of the HPIC dose rate radiation monitoring numerical anomaly and the interference effect of interpretation, and improves the auxiliary judgment ability and maintenance efficiency of ERMS radiation abnormality discovery.

Abstract:Stance detection task aims to automatically determine whether a Weibo text is in favor of the given target, against the given target, or neither. Mining the stance information about a given target is an emerging problem. Based on the success of deep learning in classifying, this study proposed a Bert-Condition-CNN model to predict the stance label. Firstly, noted that the given target may not be present in the Weibo text, so we extracted the topic phrases from Weibo corpus as the given target supplement. Then, we used Bert language model to accept the text representation vector and calculated a Condition matrix whose entries represent the relationship between Weibo text and topic phrases. Finally, a convolutional neural network was utilized to capture the stance features from Condition matrix. Experimental results on NLPCC2016 datasets demonstrate the model has achieved a sound effect of stance detection.

Abstract:In recent years, deep learning, as a hotspot of common concern in academia and industry, has made great progress and achieved remarkable achievements in computer vision, speech recognition and other fields. It is divided into two stages:training and inferencing. In practical application, the main concern is the inferencing stage. The process of deep learning inferecing is accompanied by a huge amount of computation, and more and more attention has been paid to using distributed system to improve its computing speed. However, the construction of distributed deep learning inferencing system is faced with the challenges such as rapid updating and iteration of deep learning accelerators, complex of applications and computing tasks. The information management mechanism proposed in this study is used to collect and process all kinds of information in the distributed system, and the rules of collection and processing are highly customizable and flexible. It also provides a universal RESTful API data access interface to support the flexible compatibility of various hardware and the dynamic adjustment ability of task scheduling strategy in the deep learning inferencing system. Finally, we verified the function of the mechanism through an example and analysed the experimental results.

Abstract:A railway passenger identity authentication system is designed and developed by using the application mode of "blockchain technology + asymmetric encryption + biometric identification + identity authentication" and relying on smart phone client. By the Ethereum development platform and truffle development framework, the intelligent contract of railway passenger identity authentication system is compiled and deployed. Aiming at the shortcomings of traditional railway identity authentication system, using distributed storage of passenger identity data to ease the pressure of centralized server, the system improves the security and robustness of passenger identity data. The system ensures the ownership of passenger's identity information by bioinformatics authentication, and uses asymmetric encryption technology to enhance the transparency of data under the premise of protecting passenger's privacy and realizing real-name system. The railway passenger identity authentication system based on blockchain application model can enable user identity to be stored locally and checked information abstract on the chain, realize fine-grained control of access to railway passenger identity information, guarantee the information security of railway passenger identity, and enhance the passenger experience.

Abstract:Aiming at the problems of false information flooding, lack of supervision over the whole transaction process and serious breach of contract existing in the current vehicle-freight matching platform, this study applies blockchain technology to the vehicle-freight matching platform, designs the framework and corresponding mechanism of the vehicle-freight matching platform based on public block chain, and simulates the two sides on the vehicle-freight matching platform with block chain technology with actual data. The whole process of the transaction and the conclusion of the transaction are analyzed. The results show that block chain technology can enhance the credibility of transaction information on the platform, effectively supervise the whole process of transaction execution in real time, indirectly reduce the default behavior of both sides, make up for the shortcomings faced by the vehicle and freight matching platform, and ensure the smooth progress of both sides'transactions.

Abstract:Data sharing is a powerful way to break the "Data Isolation" dilemma in the era of Big Data and how to ensure the safe sharing of data is the main problem that we face at present. Therefore, in this study, we propose a DOB framework by using Blockchain technology and ciphertext-policy attribute-based encryption. The system public key, user's attribute, ciphertext, and user secret key of the CP-ABE are stored in the Blockchain database through the smart contract and serialization method, and the access authority of the database and the registration authentication dataset are deployed to implement fine-grained sharing of data. The experimental results show that compared with the Timely CP-ABE with Blockchain scheme by Jemel et al, the proposed framework further improves the security of data sharing.

Abstract:In recent years, the Kernel Correlation Filtering algorithm (KCF algorithm) proposed by Henriques et al. shows superior performance in terms of algorithm scale, computational complexity, and algorithm performance. Based on KCF algorithm, a target tracking system based on DSP is proposed and designed in this study. In terms of hardware, this study designs and implements a complete and independent hardware platform. In terms of software, this study proposes a series of algorithm optimization methods for DSP to optimize KCF algorithm, in order to meet the requirements of important engineering indicators. The results show that the system performs well in the actual engineering environment, the highest tracking angular velocity can be 20 degrees/s, and the frame rate can be 25 fps on average, and it has high accuracy. The system provides reference for embedded applications of various algorithms in the field of computer vision.

Abstract:Wide majors partition is an important task in the management of universities. At present, Most of the partition process is completed by manual diversion, which is heavy loaded, time-consuming, easy to make mistakes, and of low efficiency. To solve those problems, it is proposed to explore the information system of wide major partition process in universities, which discussed the wide major partition admissions algorithms and professional division admission algorithms by designing the voluntary module, the large-scale division module, and the professional division module. It has solved the practical problems of diversion work in universities, and has reduced mistakes, saved labor, and saved time.

Abstract:Commodity retrieval is fundamental to the intelligent development of the e-commerce industry. This study is concerned with a fashion recognition system which bases on ZYNQ and CNN model, trains the custom network with TensorFlow and processes the weighs by using fixed-point calculations. This system applies ARM + FPGA software and hardware coordination method of the ZYNQ device to construct its framework, ARM to preprocess image by OpenCV and the CNN IP of FPGA to synchronously recognize the image. A weight-reloadable structure is implemented between ARM and FPGA, so online upgrade could be realized without modifying the FPGA hardware. The system uses fashion-minist datasets as the sample in the network training and improves the parallelism of convolution operations by increasing the acceleration engines of CNN IP. According to the experiment, the system realizes accurate and real-time identification and display for actual pictures of the e-commerce platform and its accuracy reaches 92.39%. The image processing speed can reach 1.261 ms per frame and the power consumption is only 0.53 W at 100 MHz.

Abstract:This study designs and implements a real-time intelligent parking space information query system based on deep learning object detection algorithm. The model is trained by adopting the YOLO object detection algorithm combined with a large number of vehicle and license plate images. Use the model to process the parking lot surveillance video. According to the results of the model processing and the related design of algorithm, judge the free parking space, calculate the parking time of the occupied parking space and recognize the license plate. The parking information will be received by the WeChat terminal in a schematic way, so that the drivers can obtain the parking information in real time. The system can accurately judge the parking information and provide reference for the management of urban commercial parking lot.

Abstract:This study introduces the design and the implementation of a management system of student work-study program in universities based on visual workflow technique. By separating the business management and the process management according to the characteristics of service process of the work-study program in universities, the system is realized with visual workflow technique by use of Jsp + SpringMVC + Hibernate schemes in Java EE architecture. Back-end data of the implemented system show that the designed system has significantly reduced the service approval time, enhanced the management efficiency, and thus can promote streamline operation, standardization and convenience of the work-study program management in universities.

Abstract:Aiming at the problems of the number of sites of meteorological stations, the wide distribution area, the shortage of security personnel and funds, and the high requirement of data quality, this study constructs a new cloud platform of "internet plus meteorological equipment socialization security" based on crowd sourcing. Through special link connection, it adopts the mechanism of push or request to complete the data exchange on private cloud of customer and private cloud of Tianshi Company. The integration of set operation monitoring, fault judgment, maintenance and repair, spare parts management, assessment evaluation, and mobile terminal interaction provides a new idea for the socialization of national meteorological survey guarantee.

Abstract:With the continuous development of e-commerce, the volume of campus express business has grown rapidly. The traditional express pick-up mode has low efficiency in entering and leaving the warehouse, and the labor cost in the logistics link is high. Based on this, this study proposes a communication management system based on RFID and ZigBee, which uses ZigBee for networking and RFID technology to realize automatic management of express delivery. RFID tags are attached to the packaging of each cargo, and fixed readers are set on the shelf of the station to obtain information such as the location of the goods. The user scans the express label on the exit gate channel and the QR code on the mobile phone to match, to complete the self-pickup. It improves the efficiency of the express delivery.

Abstract:Big data of petroleum industry has infinite potential and value. The application of big data and data mining technology can not only improve the industrialization level of petroleum industry, but also play a strong role in promoting the intelligent development of petroleum industry. This paper presents a new industrial knowledge mining system-Petroleum Industry Data Mining System, which is driven by Web architecture and integrates five modules of data mining. The five modules inclide data set management, pre-processing algorithm management, data mining algorithm management, data mining process management, and data result visualization. The system realizes completely self-service data extraction and data pre-processing, and completes knowledge mining process of management, data analysis, knowledge mining, and visualization of results. The flexibility of the system is greatly improved by satisfying the real-time requirements of users at different levels in different scenarios in the form of Web. Through this system, oilfield technicians can neglect the construction of large data and other complex construction processes, and better serve oilfield data modeling and analysis.

Abstract:Among all kinds of time series classification algorithm, algorithms based on local features of time series data have achieved reasonable results. However, there is still abundant space for improvements of them in time complexity and accuracy. In this study, we propose an improved algorithm based on local features. It focuses on the property of local features and put restrictions on the set of local features. On the one hand, supported by theoretical analysis, our new algorithm cuts the size of set of local features and consequently reduces the time and space complexity. On the other hand, we redefine the criteria of selecting local features so that we can select more discriminative local features.

Abstract:This article intercepts the horror audio clips from the network and movies to build terrorism audio dataset. However, the source of the horror audio is limited, whereas the convolutional neural network depends on a large amount of data. To this end, the transfer learning technology is performed in the discrimination of the terrorism audio. Firstly, pre-train the network by using the public TUT acoustic scenes dataset, and then retain the model weight and transfer the neural network to the discrimination of terrorism audio. Finally, add more layers after the fine-tune network to utilize more audio information, the structure of the added layers is similar to the residual network. The experimental results indicate that the average discriminant rate of the transfer learning method is 3.97% higher than that of the non-transfer learning method, which effectively solves the training problem caused by small audio dataset in the study of terrorism audio discrimination, and the average discriminant rate of the improved transfer learning network has increased by 1.01%, finally reaches the discriminant rate of 96.97%.

Abstract:At present, semi-supervised classification research of time series mainly focuses on univariate time series, due to the complex relationship between Multivariate Time Series (MTS) variables, there is less research on semi-supervised classification of MTS. In view of this, we proposes a semi-supervised MTS classification method based on Two-Dimensional Singular Value Decomposition (2DSVD), which first computes the eigenvectors of row-row and column-column covariance matrices, and then extracts feature matrices from MTS samples. The number of rows and columns of the feature matrix is not only lower than the original MTS sample, but also clearly considers the two-dimensional nature of the MTS sample. The experimental results on 10 MTS datasets show that the semi-supervised classification performance of this method is significantly better than the method using extended Frobenius norm, center sequence, and based on one dimensional singular value decomposition.

Abstract:Identity authentication is crucial for internet security, the widely adopted authentication technology relies on PKI system to issue certificates for servers, thus strongly depend on CA, which may face problems including key leakage and single point failure. Based on Blockchain and DNSSEC technology, this study puts forward a new model for identity authentication which can perform two-way authentication between server and client. In the mean time, improvements are made to user certificate for the management of trustable device of users thus enhance security and flexibility. This paper briefly summaries recent research in identity authentication technology at the beginning. Then it provides a thoroughly description of the new model, including the structure, work flow, and main functionality. At last, the paper analyzes the model and provides an example case.

Abstract:As a new generation of high-precision biometric recognition technology, palm vein pattern recognition technology is widely used in the field of personal identification. Effective extraction of palm vein features is essential for palm vein classification. However, because of the poor quality of the palmar vein images collected, it is necessary to enhance the palm vein image before recognition. The two-dimensional discrete fast Fourier transform (2D-FFT) is used to replace the traditional spatial convolution filtering to realize the frequency domain convolution filtering of the Gabor filter and the original image. The experimental results show that the proposed enhancement method has better enhancement effect than the traditional adaptive histogram equalization and Retinex algorithm. Compared with the traditional Gabor spatial convolution filter, it has lower computational complexity and is more suitable for real time systems.

Abstract:Currently, video analysis is usually based on video frames, but video frames usually have a lot of redundancy, so the extraction of key frames is crucial. The existing traditional manual extraction methods usually have the phenomena of missing frames, redundant frames and so on. With the development of deep learning, compared with traditional manual extraction methods, deep convolution network can greatly improve the ability of image feature extraction. Therefore, this study proposes a method to extract key frames by combining the depth feature extraction of video frame with the traditional manual feature extraction method. First, the convolutional neural network was used to extract the depth features of video frames, then the content features were extracted based on the traditional manual method, and finally the content features and depth features were fused to extract the key frames. The experimental results show that the proposed method has better performance than the previous key frame extraction method.

Abstract:Image semantic segmentation methods based on deep convolutional neural network requires a large number of pixel-level annotation training data, but the labeling process is time-consuming and laborious. In this study, a semi-supervised image semantic segmentation method with encoder-decoder based on generative adversarial networks is proposed, in which the encoder-decoder as the generator. The entire network is trained by coupling the standard multi-class cross entropy loss with the adversarial loss. In order to make full use of the rich semantic information contained in the shallow layers, this study puts the features of multi-scales in the encoder into the classifier, and fuses the obtained classification results with different granularities to optimize the object boundaries. In addition, the discriminator enables semi-supervised learning by discovering the trusted regions in the unlabeled data segmentation results to provide additional supervisory signals. Experiments on PASCAL VOC 2012 and Cityscapes show that the proposed method is superior to the existing semi-supervised image semantic segmentation methods.

Abstract:Lane-crossing detection of vehicles is an important part of intelligent transportation system. To tackle this issue, we proposed a lane-crossing detection method of vehicles with in-vehicle image. First, we use synthesizing-data method to build a rich and varied lane-crossing detection dataset. Then, we use image semantic segmentation to detect vehicle and lane lines, and then we estimate wheels positions of vehicle. Finally, we contrast wheels positions with lane lines positions to judge whether there is lane-crossing behavior. Experiment results show that combined with semantic segmentation model, we achieve an average precision of 88.7% for lane-crossing detection, and the average detection time is 35 ms, which means that the proposed method has certain practical application value.

Abstract:Focused on the low-light images obtained from dynamic range, illumination condition, image acquisition equipment, etc., a feature enhancement derivative fusion algorithm based on luminance evaluation technology was proposed to achieve contrast adjustment and feature enhancement of the low-light images. Firstly, the brightness evaluation technique was used to optimize the brightness of the low-light image to obtain the exposure ratio map. Then, combining exposure ratio map and improved chi-square distribution function model, two derivatives with enhanced features were obtained for fusion. Finally, the fusion image was obtained by using the improved derivative fusion algorithm. The experimental results indicate that the proposed algorithm achieves the better results including brightness order error, visual information fidelity and image mutual information, improves the image contrast while preserving the well-exposed region, and it can recover the edge and texture details of the low-luminance region.

Abstract:For low-level computer vision tasks, image de-raining has always been a hot issue. However, due to the uneven density of rain lines in the image, it is a very challenging issue to remove rain from a single image. Attention to the target image often requires attention to two parts:the overall structure of the image and the details of the image. In this regard, a novel multi-stream feature fusion convolutional neural network algorithm is proposed. It presents superior performance through various network frameworks. The network algorithm uses three branch networks to extract complex multi-directional rain line features, concat features and combines with the original image to remove rain. The detail enhanced network can obtain high quality without rain. The de-raining performance under the synthesized dataset and the real rain dataset indicates that the proposed algorithm can preserve more details while removing the rain line than the existing deep learning-based rain removal algorithm, and it can keep the high quality of the picture.

Abstract:With the development of Virtual Reality (VR) technology and the increasing demand for human-computer interaction performance and experience, gesture recognition is one of the important technologies affecting the interaction in VR, and its accuracy needs to be improved. Aiming at the problem that the current gesture recognition method performs poorly in some similar gesture recognition, a multi-feature dynamic gesture recognition method is proposed. Firstly, this method uses Leap Motion to track the dynamic gestures to acquire data, then adds the displacement vector angle and the inflection point judgment in the feature extraction process, after that performs the training of the dynamic gesture Hidden Markov Model (HMM). Finally, the recognition is carried out according to the matching ratio of the gesture to be tested and the model. It is concluded from the experimental results that the multi-feature recognition method can improve the recognition rate of similar gestures.

Abstract:The detection of traffic sign is the crucial technology of traffic sign recognition system. A method of traffic sign detection based on image color and shape is proposed. Firstly, the image is pre-processed by gray stretching and noise filtering, and then the color image is segmented by improved K-means clustering algorithm. Finally, the shape detection technology based on Hough transform is used to locate the special shape of traffic signs, so as to realize the detection of traffic signs. The experimental results show that the average accuracy of the detection results under various complex background conditions is 93.0%, which is better than other algorithms under the same conditions and has high real-time performance.

Abstract:In the indoor positioning, the traditional RFID positioning method cannot accurately estimate the current path loss coefficient with the change of indoor environment due to its simple method. It has disadvantages such as large environmental impact, low positioning accuracy, and poor real-time performance. In order to solve the above problems, this study puts forward a kind of indoor location algorithm based on dual neural network model, and establishes the BP network and the network within DNN dual neural network model. Then, it preprocesses the collected RSSI signal and inputs the preprocessed signal value to BP network model, outputs path loss coefficient n, and then received signal strength value RSSI and through the BP model to get the path loss coefficient of n as input, input to the network within DNN model, and get the precise positioning of the labels under test coordinates. Experiments show that compared with the traditional indoor positioning algorithm based on RSSI and ANN model, this algorithm effectively improves the positioning accuracy and real-time performance.

Abstract:About radar emitter signal identification research, the artificially extracted features have relatively physical characterization, but there are still redundant features and noise features. Through the deep neural network, the deeper expression of the signal can be obtained, but its characteristics are difficult to explain. Combining the physical characteristics of artificial features and the strong learning ability of deep learning, this study proposes to apply a deep feature selection network to radar signal recognition technology. DFS adds a sparse one-to-one layer between the input layer and the first hidden layer to obtain the corresponding weight value of each feature from the classification correlation metric, uses these weight values to enhance the input of sensitive features and weaken the input of redundant features, and improves classification accuracy. Firstly, the complexity features, Cscade Connection features of ridge-frequency, and information entropy features are extracted from the radar signals, and merged into the original feature set. The DFS is used for learning training to achieve the feature selection at the input level. The above approaches were used to identify the 5 different types of radar emitter signals, obtained good classification. The results verify the effectiveness of the approach.

Abstract:At present, the research and application of video technology mainly focus on the processing of video images by image processing and pattern recognition. The video capture is started from the capture time point, thus the information contained in the captured video lacks integrity, that affects further analysis and processing. Aiming at this problem, design and implementation of a video capture scheme are proposed based on Android system by defining the time window and using the cache technology. The key is to provide video for a certain period of time from before to after the capture time point. The test results show that the capture scheme has a small time error and occupies less system resources, and has been applied to vehicle electronic systems which are mass-produced and put into the market, providing an important basis for traffic accident liability identification.

Abstract:Documents often contain horizontal lines, hand lines, etc., which are used for various special functions. When these documents are stored in computers by scanning or the like and need to be further recognized and processed into text codes, these lines become interference factors of OCR, thus the recognition rate of document content is decreased. This study proposes a new document interference line removal algorithm, which first binarizes the document image, and the binarization process takes into account the effects of uneven illumination; then the foreground is refined into single pixels, reducing the thickness of the lines. The effect is then calculated by an improved greedy algorithm to calculate the weights of the horizontal and vertical line segments, and the line segment with higher weight is determined as the interference line; finally, the distance of each foreground pixel in the image is determined by the distance from the interference line. Thereby obtaining a complete document recovery map. The simulation results show that the proposed algorithm can effectively remove the interference lines, especially in the case of interference lines and text adhesion, and remove the interference lines while affecting the quality of document images less, and has a higher computing speed and better removal effect. The removal effect provides a good basis for further OCR recognition of images.

Abstract:Mobile app advertising is a kind of initiative advertising in the Internet advertising market. It can analyze users' interests and hobbies to target their interest advertisements, and then it improves their experience and brings huge benefits to advertising platforms and advertisers. Therefore, predicting the conversion rate of mobile app ads has become a significant research direction. Based on logistic regression and two gradient boosting tree models, this study proposes two integration models, SXL and BLLX, using the idea of stacking and averaging. It solves the problem that traditional prediction models have limited capabilities and cannot accurately predict the conversion rate. The experimental results on the dataset of Tencent 2017 social advertising competition show that the SXL and BLLX can effectively improve the prediction results of advertising conversion rate.

Abstract:In the process of emergency rescue, due to the lack of optimization of collaboration mode between resource entities such as emergency teams and the lack of high matching degree of dispatch, this study proposes a method based on case base and plan base to calculate the skill contribution degree of emergency execution entities, the relationship strength and collaboration degree between entities, and to mine the high collaboration mode among emergency entities by using the high collaboration algorithm of emergency entity trajectory mining. The completeness and matching degree of high emergency team scheduling management provide basic data. At the same time, based on the activity trajectory and activity continuity of entities, high continuity activities can be excavated, which lays the foundation for decision-making of emergency response sequence in rescue. The experimental results show that the coordination ability of the same and different emergency entities varies with the number of collaborations, which deviates greatly from the results of previous empirical decision-making. It provides important data support for the rational dispatch of emergency teams and the formulation of efficient emergency plans, and has practical value.

Abstract:The traditional workflow framework enhances the design orchestration and manageability of process tasks, but there are problems such as the lack of runtime expansion of the process engine, insufficient flexibility of task participants, and large change costs of complex process forms. This paper presents a containerized flexible microservices process development framework, combining domain-driven design, container-based microservices, dynamic organization key points and other technical mechanisms to improve workflow implementation efficiency, reduce development costs, and better solve the above problems. After the experimental analysis, the flexibility is enhanced and the cost is smaller, and the effectiveness of the framework and method is verified in practical projects.

Abstract:Photovoltaic modules inevitably produce various defects in daily operation, and hot spot defects are one of them. The existing research mainly focuses on the defects of photovoltaic modules in the production process, and there is few research on the defect detection algorithms generated by PV modules in daily operation, and there are problems such as poor generalization ability and insufficient accuracy. Based on the original faster RCNN, this study combines image preprocessing, migration learning, improved feature extraction network model, and improved anchor frame selection scheme to obtain hot spot defect detection model. Experiments show that the average detection accuracy of the self-made test set using this model is 97.34%, which is 4.51% higher than the original faster RCNN.

Abstract:Micro-service architecture has gradually become the mainstream architecture for building complex business systems, but at the same time, the system architecture has become more complex, and the quality of service assurance becomes a key issue. This study focuses on the quality of service assurance under the micro-service architecture, and solves various problems in the process of service invocation by fusing, degrading and multi-channel invocation. Based on real-time monitoring data of business and equipment, a calculation model of service invocation parameters is proposed, which can refine the calculation of service invocation parameters and support online dynamic parameter adjustment. Through the elegant shutdown and flow preheating method, the quality jitter problem in the process of service restart is solved. The method and model proposed in this study have been fully validated in practical application. The response time and failure rate are greatly reduced, which makes the system run more stably and reliably under the micro-service architecture.