Abstract:The diversity of crowd scale in reality is a great challenge to crowd counting algorithms. Therefore, a novel crowd counting algorithm based on scale fusion is proposed in this study. Firstly, the algorithm for density map generation is optimized. Multiple head detectors are used to obtain part of the head scales of the sparse crowd, and RBF interpolation is employed to complete this part of the density map. As to the dense part of crowd, the traditional distance self-adaptive algorithm is adopted to generate a more accurate density map. Secondly, the regression neural network of the density map is designed with a mobile inverted bottleneck convolution module, and a dilated convolution module is added to facilitate the extraction of head edge features. Finally, the loss function of the regression neural network is optimized by distinguishing the crowd area from the non-crowd area. In the experiment part, the algorithm is compared with other similar algorithms on multiple datasets, and the results show that the proposed method can significantly improve the accuracy of crowd counting.

Abstract:The device fault generally starts from a minor one and gradually develops to the loss of working capacity of the whole set. Detection in case of a minor fault can recover the unnecessary loss. Therefore, this study proposes a method to evaluate device health status on the basis of the Weighted Mahalanobis Distance (WMD) and the Device Status Index (DSI). Based on an improved Mahalanobis-Taguchi system, the method constructs a stable reference space for the characteristic parameters during the effective operation of the device. It selects the characteristics and calculates the WMD according to the device fault sensitivity, eliminating the interference of characteristic correlation. Then Box-Cox transformation is used to determine the threshold value of the DSI to build a health status model of the complex heavy device, and the model is verified by experiments. The WMD values of the normal samples are all below the fault threshold, and nearly 98.6% of the sample values are within the warning signs. The proposed method can provide data support for maintenance and management of complex heavy devices, thereby facilitating industrial production.

Abstract:As the intelligence level grows, a large amount of new knowledge is generated all the time, and knowledge graph has gradually become one of the tools for knowledge management. However, the existing knowledge graph still has some problems, such as missing attributes, sparse relations, and massive noisy information, which leads to poor graph quality and is easy to affect various tasks in the field of natural language processing. As a research hotspot, the knowledge reasoning technology oriented to the knowledge graph is the main method to solve this problem. It improves the information of the knowledge graph by simulating the human reasoning process, with a good performance in many applications. Taking the knowledge graph as the pointcut, this study classifies and explains the knowledge reasoning technology by categories and elaborates on several application tasks of the technology, such as intelligent question-answering and the recommendation system. Finally, it forecasts the main research directions in the future and puts forward several research ideas.

Abstract:Due to the shaking of a handheld camera or the movement of targets, the video image data is subject to motion blur, which reduces the image quality of human perception. With regard to the problem, from how to obtain clear images from the original process to how to obtain clear images efficiently, a new model for real-time video image deblurring based on the lightweight Generative Adversarial Network (GAN) is proposed in this study. The model defines PatchGAN as a discriminant network and sets up a dual-scale discriminator for global images and local features on the basis of it; the generation network takes lightweight MobileNetV3 as the backbone network and introduces a feature pyramid for feature extraction to solve the problem of low utilization of feature information in the discrimination network and low inference efficiency of the generation network. This model uses an end-to-end approach to efficiently deblur the video image. After experiments on the GoPro and Kohler datasets, the results show that the sharp image deblurred by this model has a high peak signal-to-noise ratio and great structural similarity, and the inference speed reaches 1.7–127 times faster than that of other models.

Abstract:The next Point-Of-Interest (POI) recommendation is one of the most important services of the Location-Based Social Network (LBSN). It can not only help users find the destination which they are interested in, but also improve the potential income of business providers. Existing algorithms have employed user behavior sequences and the POI information for recommendation, but none of them fully utilize POI side information, thereby failing to ease the problems of cold start and sparse data. In light of the above analysis, this study proposed a POI recommendation system, Graph Embedding-Gated Recurrent Unit (GE-GRU). Firstly, GE-GRU relies on Graph Embedding (GE) to integrate the POI itself with its side information to get the POI embedding that contains deep information. Then, the POI embedding is input into the GRU-based neural network to model recent user preferences to acquire user embedding. Finally, according to the POI rank list, the next POI can be recommended. Experiments are conducted on a real dataset, Foursquare, which contains more than 480 000 check-ins, and Accuracy@k is adopted for evaluation. The results show that, compared with GRU and Long Short-Term Memory (LSTM), GE-GRU has 3% and 7% improvement on Accuracy@10, respectively.

Abstract:Streamline rendering has long remained as one of the most common techniques for flow visualization. The streamline is an effective sparse representation of the flow field, which can capture the flow behavior, but generating streamline needs long-term particle tracing and massive integral operations. Large-scale flow visualization takes considerable computation time, and the parallel computing algorithm and high-performance equipment are needed. In this study, a high-resolution streamline generation algorithm based on deep learning is designed. The initial sparse low-resolution streamline is quickly mapped into the dense high-resolution streamline to provide reliable streamline visualization results in a short time. On this basis, an interactive real-time flow visualization system is developed, which is capable of flow-field feature detection, attribute correlation analysis, information theory analysis, etc. It can help users quickly understand the flow field data and find their areas of interest for post-hoc analysis, avoiding redundant data and enhancing work efficiency. In addition, it can meet the users’ needs for multi-dimensional correlation analysis of flow field structures, features, and attributes.

Abstract:It is difficult to count the discipline information stored in a scattered form. With regard to this problem, based on the domain ontology model of computer discipline, the computer discipline knowledge graph in universities is constructed by integrating the multi-source and heterogeneous data. First, domain knowledge is acquired from relevant websites and existing documents through Web crawlers and other tools, and the data are cleaned on the basis of the BERT model. Then, Word2Vec is used to judge the similarity between the research directions of characters, so as to solve the problem about entity alignment. Finally, the data are imported into the Neo4j graph database to realize the storage of knowledge. According to the knowledge graph, the visualization system of computer discipline is established, which can fulfil information retrieval, graphic display, and other functions and realize quick query and resource statistics of computer discipline data. It is expected to facilitate the follow-up discipline evaluation work and make it more efficient.

Abstract:With the rapid development of e-commerce platforms, the logistics industry is at a high rate of growth. The access logs of the logistics service platform can reflect user behavior, so it is very important to tap the hidden information to help the logistics service platform optimize the business. At present, higher real-time requirements are imposed on large-scale log data processing. This study comprehensively considers a variety of stream processing frameworks capable of real-time computing, large-scale storage databases, log collection tools, etc. It chooses Flume and Kafka as the log collection tools and message queues and uses Flink and HBase for real-time calculation of streaming data and large-scale data storage. At the same time, the functions including data deduplication, abnormal alarms, fault tolerance strategy, and load scheduling are designed for the platform. Experimental tests have proved that this processing platform can efficiently process log data of the logistics service platform, with innovative ideas and practical value.

Abstract:The phenomenon of impaired concentration is common among adolescents, but the existing systems for attention detection and training are equipped with simplistic functions. This study develops an attention detection and training system based on EEG signals for adolescents. In light of few classifications and low accuracy in attention detection, we divide attention into five categories and propose an attention detection method based on the random forest model for higher detection accuracy which arrives at 76.17%. With regard to the deficiencies in unsatisfied effect of attention training, we design three serious game training modes for adolescents in terms of sustained attention, selective attention, and focus attention for the first time using the EEG-based closed-loop biofeedback technology. Meanwhile, to verify the effectiveness of the attention training mode, we define four indicators to conduct experiments with the self-control method while excluding the influence of familiarity with the game on subjects.

Abstract:Xinjiang is an arid region with a typical dry climate. The economic foundation is relatively weak and the degree of water conservancy intelligence is not high. The utilization rate of water resources for agricultural irrigation is only about 40%. Jeminay County is the most typical water-deficient county in Xinjiang. The snow line of Muz Taw Glacier, the only water source in its territory, has constantly retreating due to global warming, leaving an urgent situation. Traditional water allocation and management methods are relatively backward, with high labor intensity, low efficiency, and poor benefit. It can no longer solve the current dilemma about water resource management in Jeminay County, nor adapt to the current economic situation. Modern smart water conservancy has high efficiency by leveraging information technology, network technology, big data, and artificial intelligence. Based on the smart water conservancy management platform of Jeminay County in Xinjiang, this study demonstrates the construction details of the smart information integration platform from seven aspects and provides strong support for promoting the construction of a smart water conservancy system.

Abstract:In the current information environment, unstructured text is an important part of information. How to quickly extract the required information from text data and provide users with an efficient way to acquire information has become an urgent problem to be solved in the current information service field. Based on semantic retrieval and an extraction-type document reading comprehension model, this work studies how to effectively extract required answers from large document libraries according to users’ questions and constructs an information service question answering system based on document libraries. In the current environment with massive information, it is of great significance to improve the efficiency of users’ information acquisition. Experiments show that the system can quickly and accurately locate the answers to the users’ questions and help them get the required information rapidly.

Abstract:During classes, students’ mastery of knowledge points is reflected in their facial expressions. Teachers usually judge their degrees of understanding through the expressions and then adjust the teaching schedule. However, at least 30 students are in a class, and it is impossible for the teachers to take care of each student at all times in a classroom. As a result, teachers can not fully understand each student’s mastery of knowledge points, thus affecting the quality of teaching. To solve this problem, this study introduces a classroom teaching feedback system based on facial expression recognition, which can analyze the facial expression of each student in the classroom and demonstrate their mastery of knowledge points from the expressions. On that basis, it can help teachers understand the real-time learning effect of each student in the classroom, thereby improving teaching quality.

Abstract:With the popularity of video surveillance, action analysis technology based on artificial intelligence is playing an increasingly important role in the field of intelligent surveillance. Most of the existing algorithms depend on an optical flow network or a 3D network to obtain the time information of actions. However, the optical flow network and the general 3D network require a large amount of computation, and the computational efficiency is low when classification and localization are carried out simultaneously. To solve this problem, this study builds a dualflow framework capable of spatial localization and classification and follows the idea of SVD to decompose the 3D convolution kernel in the branch of the 3D network, thus reducing the 3D network parameters. In addition, the dynamic programming algorithm is employed to efficiently search the optimal action tubes, and the mixup algorithm is used to expand the data sets during training, thereby enhancing the training results. Finally, experimental verification is carried out on UCF101-24 and J-HMDB-21, two widely used data sets for action localization. Compared with the baseline algorithm, the Frame-mAP of the two data sets is improved by 7.1% and 4.8%, and the Video-mAP of J-HMDB-21 under different IoUs is enhanced by 5.2% and 4.8%. Experimental results show that the proposed algorithm can substantially improve the ability of action localization, with better results compared with other algorithms.

Abstract:Soybeans include many varieties (cultivars) and their cultivars have subtle differences in leaf patterns, which makes it tough to distinguish them from leaf features. Great progress has been made in using leaf image patterns for plant species recognition. However, as a general fine-grained pattern recognition problem, soybean cultivar recognition has not yet received considerable attention. Traditional hand-operated leaf image analysis is limited to capture the subtle differences of leaf features among different cultivars. In this study, we attempt to use deep learning to harvest discriminatory leaf features for soybean cultivar recognition. A novel deep learning model, Transformation Attention Network (TAN), is proposed in this work. It first extracts fine-grained leaf features via the attention mechanism and then rectifies the leaf posture by affine transformations. We construct a soybean leaf cultivar dataset which consists of 240 soybean cultivars, with 10 samples per cultivar, to examine the availability of cultivar information in leaf patterns and validate the effectiveness of the proposed deep learning model for soybean cultivar recognition. The experimental results confirm the effectiveness of the leaf image patterns in distinguishing cultivars and demonstrate the better performance of the proposed method than that of the state-of-the-art hand-operated methods and deep learning methods in soybean cultivar recognition.

Abstract:In Software-Defined Networking (SDN), inconsistent updates are frequent when the controller needs to update the flow table entries in multiple switches due to the change of configuration policy. The internal reason of this phenomenon is that the controller updates all switches asynchronously. Update delay leads to the logical inconsistency of the network state, affecting the correct forwarding of data messages. With regard to poor generality of application scenarios and prolonged update time of the classification and sequence based update scheme and high computational complexity of the optimal update scheme, this study proposes a Categorical Search based loop-free Consistent Update scheme (CSCU) for flow tables. In this scheme, a switch classification model is developed on the basis of classification, and a loop search optimization model is built according to the idea of node dependence. Those contributions achieve the consistent update with short update delay and high update efficiency. The simulation results show that the proposed scheme has promising applicability with lower node operation complexity as well as fewer update rounds, which can markedly improve the update performance in terms of lower computational complexity.

Abstract:To eliminate the shortcomings of low precision and slow convergence in the sales forecasting based on the traditional BP Neural Network (BPNN), this study proposes a new model based on an Improved Immune Genetic Algorithm (IIGA) optimized BP neural network. IIGA presents a new way of population initialization, a regulatory mechanism of antibody concentration, and a design method of adaptive crossover operators and mutation operators. Therefore, the convergence ability and global search ability of IIGA are greatly improved. In addition, IIGA can optimize the initial weights and thresholds of the BP neural network and overcome the drawbacks of output instability of the BP neural network and proneness to local minimum induced by the randomness of network parameters. With the past records of sales figures in a steel enterprise as an example, the BP, IGA-BP and IIGA-BP neural network forecasting models are built with Matlab for simulation comparison. The experiments demonstrate that the precision of the IIGA-BP model is 23.82% higher than that of the BP model and 22.02% higher than that of the IGA-BP model. The IIGA-BP model possesses better generalization about steel sales forecasting and more stable forecasting, with errors basically in the range of -0.25 to 0.25, and its forcasting precision is dramatically improved. The proposed model provides a more effective method for sales forecasting in enterprises.

Abstract:Traditional Particle Swarm Optimization (PSO) is likely to converge to local optima when applied to multimodal problems, with low search efficiency. In this study, a novel multi-swarm PSO algorithm based on swarm relations and repulsion factors is proposed, called Swarm-Relation-Based PSO (SRB-PSO). Three swarm relations, including dominance, equivalence, and weakness, are defined according to the search results. The search diversity is guaranteed by introducing repulsion factors among equivalent populations and the search efficiency is increased by dominance and weakness relations. Thus, the global search ability of the algorithm is enhanced and the solution quality is improved. The new algorithm and several other versions of PSO are compared on a set of benchmark functions. The results show that the algorithm proposed in this study can well maintain the particle diversity and has outstanding global search ability. The proposed algorithm outperforms the other algorithms when solving multimodal problems.

Abstract:At present, the logo recognition technology in China is being rapidly developed, which is embodied in processing accuracy, reproducibility, flexibility, applicability, and information compression. However, the development of this technology is still limited by actual demands. The deep learning model has heavy computation and is difficult to run on lightweight embedded devices. There are many and complex noises in industrial production, which affect the recognition accuracy. To solve the above problems, this study proposes a logo recognition technology based on the convolutional neural network. An improved Canny edge detection algorithm is used to enhance the robustness in edge information extraction, and signs are accurately extracted in a high-noise environment. In addition, to further improve the recognition accuracy, in the combination of Convolutional Neural Network (CNN) and ellipse fitting, this study combines the model recognition and ellipse fitting results to determine the recognition accuracy. This method improves the recognition accuracy while increasing a small amount of calculation.

Abstract:It is a classic problem to judge whether there is a path between two vertices in a directed graph. For some practical applications such as routing and graph analysis, it is required to find whether there is a reachable path with limited hops, which is a variant of reachability query in graphs. For the query algorithm with limited hops on a large graph, it is necessary to balance the time and space efficiency of large-graph query and optimize the algorithm with the characteristics of limited hops. The common reachability query algorithm takes up too much space for small-degree vertex index entries, which leads to serious space waste. Therefore, we propose a hop-limited 2-hop partial index method, which combines an improved index method with local search to achieve hop-limited effective reachability query. The experimental results show that, compared with the existing algorithms, the proposed algorithm can save 32% index space and slightly increase query time on the Orkut social network dataset. Thus, the proposed algorithm can calculate the hop-limited reachability problem of larger graphs.

Abstract:Heart rate is an important basis for reflecting the status of the cardiovascular system in humans. Video-based non-contact heart rate detection has been widely applied due to its advantages of strong scene adaptability and low cost. However, this method is susceptible to noise interference such as illumination change and target movement. To solve this problem, this study proposes a method based on the decomposition of vibration phase signals to extract the video heart rate on the wrist skin. Its core idea is to find the band range of pulse signals by designing a direction-selective complex steerable pyramid. The signal-to-noise ratio is used to select the pulse signals of interest and the robust principal component analysis is employed to isolate pulse signals from the mixed signals. Finally, the heart rate of noise-resistant pulse signals is detected. In this study, the data set of heart rate detection is collected, and the output from a pulse detector is taken as the true values to verify the method. The accuracy is 97.80% in the interference scene, which is over 5% higher than that of the three mainstream methods.

Abstract:The majority voting mechanism based on heterogeneous redundant architecture realizes the fault tolerance of mimic defense systems. In the Mimic Common Operating Environment (MCOE), the mechanism is achieved after the external voting module performs the majority voting on the response data of heterogeneous executors. To improve the voting mechanism and improve the voting speed, this study proposes a majority voting mechanism based on historical performance security and heterogeneous confidence and a parallel clustering algorithm. The improved majority voting mechanism can effectively tackle the problems of the simple majority voting mechanism that cannot produce voting results and ignore the security of the executors themselves and their correlation. The parallel clustering algorithm solves the problem of idle data in the voting process and significantly improves the voting speed. In addition, this study designs a specific processing flow for structured and unstructured (images, audios, videos) data to ensure the reliability of data comparison.

Abstract:At present, the computer aided diagnosis technology based on medical imaging is at a stage of rapid development, but limited by the medical imaging data size, the modeling method based on deep learning cannot explore more complex models. Starting from the data augmentation method for medical CT images, this article summarizes the imaging characteristics of medical lesion images, classifies the existing methods for lesion detection and segmentation tasks, and expounds on the current difficulties in medical image detection and segmentation. It summarizes the related technologies of medical lesion detection, the data augmentation methods, and the lesion detection methods based on the Generative Adversarial Network (GAN). Finally, the data augmentation methods based on deep learning in the medical field, including GAN, pix2pixGAN, and CycleGAN models, are comparatively analyzed, and the future development trend of medical image analysis is prospected.

Abstract:One of the goals of “Industry 4.0” is to build traditional factories into intelligent ones. With the emergence of intelligent factories, traditional network security cannot meet the needs of enterprises and users. In view of the security risks in intelligent factories and their products, such as proneness to privacy disclosure, this study proposes an ultra-lightweight authentication scheme suitable for intelligent factory systems by combining the RFID technology with the Blockchain technology. The scheme combines the classic RFID technology with the emerging Blockchain technology to reduce the amount of calculation under the condition of ensuring security. It meets the security needs of users with the mechanism of Blockchain decentralization. Based on the bidirectional authentication mechanism in RFID, it can resist common attacks and has high security and computing advantages.

Abstract:The variation trend in pollutant concentration is of great significance for environmental monitoring. At present, the output of various feedforward neural network prediction models is only related to the current input, so it is impossible to study the dependence of pollutant data before and after. Multiple pollutants have the same emission source, and there is a potential correlation between pollutants. The change in one pollutant may reflect that in another pollutant, so the influence of other sensitive parameters should be considered in the prediction. To solve the above two problems, this study proposes a regional key pollutant concentration prediction method based on sensitive parameter discovery. The association rule algorithm and multiple regression analysis are used to mine the sensitive parameters of each pollutant, and a multivariable LSTM prediction model is constructed. The pollutants to be predicted and their sensitive parameters are taken as the characteristic variables of the model to predict the pollutant concentration. The experimental results show that the proposed method can reliably predict the variation trend in pollutant concentration and it performs better than the LSTM model without relation discovery.

Abstract:Regarding the problems of excessive divisions and high computational complexity in the current measurement division method of extended target tracking, this study combines the Clustering by Fast Search and Find of Density Peaks (CFSFDP) with the Box-Cardinalized Probability Hypothesis Density (Box-CPHD) filter to propose a Box-CPHD extended target filtering algorithm based on CFSFDP. The algorithm applies CFSFDP to measurement division and it can clearly divide the interval measurement and remove the clutter measurement with the difference in the measurement information density. Then, the Box-CPHD filter is used for prediction update and target state estimation. The simulation experiment shows that in comparison with the classic distance division method, CFSFDP is employed in the measurement preprocessing of the Box-CPHD extended target algorithm. CFSFDP significantly reduces the running time while achieving the same effect, and in the high-clutter environment after clutter removal, the change in clutter only affects the calculation time of distance division but no longer affects the CFSFDP division. The processing of measurement information with CFSFDP can greatly improve the operating efficiency and the real-time performance of the algorithm. After clutter removal, the accuracy of target position estimation is improved to a certain extent.

Abstract:The label propagation algorithm, a commonly used community discovery method, has approximately linear time complexity but randomness and instability. To solve the problems of low accuracy and poor stability of the label propagation algorithm, this study proposes an improved Label Propagation Algorithm based on node Importance and Similarity (LPA_IS). First, based on node importance, a method is proposed to obtain the seed node set and the algorithm update sequence. Second, a method is proposed with node importance and similarity to calculate the comprehensive influence of labels. Any node updates its own label according to the comprehensive influence of its neighbor labels. Experiments on real networks and synthetic networks have shown that compared with five typical label propagation algorithms, LPA_IS can improve the accuracy and stability to a certain extent and reduce the iterations.

Abstract:To solve the difficulty in detecting the defects in fabrics of unknown styles in the automatic fabric defect detection algorithm, this study proposes a fabric defect detection method based on feature residuals. First, the defect residuals of the defective and template fabric images are fused with that of the normal unlabeled fabric image to generate a new defective fabric sample. Then, the improved feature extraction network uses the shared weight method to extract features from the defective and template fabrics and calculate the feature residuals. Finally, the ROIAlign method is used to mix the global context information and the region of interest for feature fusion. The fused features are subject to defect classification and location return. Experiments are separately conducted on two test sets containing and not containing fabrics of unknown styles. The results show that the improved algorithm can better eliminate the influence of fabric styles on the detection results. The accuracy is greatly improved in the test set that doesnot include unknown styles. In the test set containing unknown styles, defect detection maintains high accuracy. Compared with that of the general algorithm before the improvement, the final scores have increased by 15.4% and 16.2%, respectively.

Abstract:Fine-grained image classification is challenging due to the difficulty in the effective learning of discriminative objects in images. Therefore, this study proposes a weakly supervised fine-grained image classification algorithm based on the attention mechanism. This algorithm can accurately locate and identify the semantically sensitive features in fine-grained images. First, on the basis of the classic convolutional neural network, the overall information of an object can be expressed by the linear fusion of features. Then, the discriminative details of the features are further extracted through the visual attention mechanism to obtain a more complete fine-grained feature expression. The proposed algorithm combines linear fusion with the attention mechanism and it can be regarded as a network model of multi-network-branch cooperative training and joint optimization. Thus, the network model can better express the overall and local information. Experiments on three publicly available fine-grained identification datasets show that the proposed method is superior to the baseline method and achieves the advanced classification level.

Abstract:In the A* algorithm, path finding is slow and the generated path has redundant turning points. For these reasons, an A* algorithm with variable step sizes based on the steady-state steering model of vehicles is proposed. Firstly, the search step size of the A* algorithm is adjusted by setting sub-targets to reduce path finding time. Secondly, local replanning is performed according to the kinematic constraints on vehicle steering at the turning points of the global path. Thus, a smooth path of easy tracking is obtained. In addition, considering the actual width of an Unmanned Ground Vehicle (UGV), the improved algorithm also introduces an obstacle extension strategy to make the planned path meet the actual engineering application. Finally, the proposed algorithm is proved effective. A comparison between this algorithm and three path finding algorithms shows that the improved algorithm has obvious advantages over the other three algorithms, including shorter path finding time, smoother paths, and safe distance from obstacles being maintained.

Abstract:A new co-evolutionary genetic algorithm is proposed. Based on the coevolution idea, the algorithm divides the population into groups. Each group adopts different crossover and mutation strategies according to the individual situation and difference in its own group. To prevent prematurity, this algorithm only employs the adaptive strategy to dynamically adjust the mutation factor when the catastrophic condition is not triggered. When the catastrophic condition is triggered, with the adaptive strategy applied, the catastrophe mechanism is introduced to generate some new individuals to jump out of the local optimum. The results of function optimization show the effectiveness of the algorithm. The algorithm is used to deal with flow shop scheduling with the optimization objective of minimizing the maximum completion time. The results show that the algorithm is superior to the traditional genetic algorithm in convergence speed and accuracy of optimization results and performs well in solving the shop scheduling problems.

Abstract:Regarding the low localization accuracy and coverage of the Approximate Point-In-Triangulation Test (APIT) algorithm, this study proposes a hybrid localization algorithm based on the APIT and genetic algorithms for Wireless Sensor Network (WSN). This algorithm improves localization accuracy by the APIT algorithm optimized with a comparison of segmentation methods and enhances localization coverage by the genetic algorithm. Simulation results show that in comparison with the APIT algorithm, the localization accuracy and coverage of the proposed algorithm are respectively increased by 21.62% and 4.87%.

Abstract:Histogram of Oriented Gradients (HOG) feature extraction has a slow speed and is prone to the omission of detailed features in pedestrian detection. To tackle these problems, this study proposes a novel pedestrian detection algorithm based on Gabor feature combined with fast HOG feature. Specifically, the input image is first subjected to wavelet transform and the HOG feature of the image is quickly extracted using the idea of integral image and the principal component analysis algorithm. Then the fast HOG feature is fused with the Gabor feature obtained after Gabor wavelet transform. Finally, the hybrid features are used to train the classifier for effective pedestrian detection. The experimental results on the test set show that the detection accuracy of the hybrid feature extraction method is up to 7.37% higher than that of the single feature extraction method when the same classifier is used. Therefore, the proposed algorithm can effectively improve the accuracy of pedestrian detection.

Abstract:In the era of big data, e-commerce platforms have accumulated a large number of user behavior data, such as browsing, clicking, placing orders and adding commodities to shopping carts. How to use machine learning algorithms to explore the consumer preferences and habits behind big data has become a new research hotspot. This study mainly improves the user purchase prediction from two aspects: feature engineering and model building. After the deep understanding of e-commerce knowledge, we have constructed 115 features with statistical knowledge and data from many aspects such as users, commodities and comments. Moreover, a two-layer fusion model is designed. The first layer uses XGBoost, CatBoost, and logistic regression as the base classifiers which predict user purchase behaviors from different perspectives. The second layer employs a weighted average method to fuse the prediction results of the base class model, and its weight is generated by linear classifier learning. The experimental results show that the F1 score of the fusion model is higher than that of the individual classifier, and many times of experiments prove that the fusion model has high stability compared with the individual classifier.

Abstract:To tackle the problems of large power consumption and intricate design of the automatic control system for air conditioning in a telecommunication room, this study proposes an energy-saving control method based on the Dueling-DQN algorithm and rule constraint for mechanical control system design. With the ability to learn modeling adaptively according to the environments of different computer rooms, this method can save the power consumption of air conditioning while ensuring the indoor temperature in the specified range. Moreover, according to the actual application scenarios of computer rooms, the states, actions and reward functions of the energy-saving control algorithm are designed. Besides, a deep reinforcement learning algorithm Dueling-DQN is used to improve the model expression ability and learning efficiency. The results of actual verification in telecommunication rooms show that the control method can save energy by 18.3% compared with the air conditioning at default parameters. It can be easily extended to machines in different environmental scenarios to provide solutions for energy conservation and emission reduction of telecommuni-cation rooms.

Abstract:Medical spine CT images have low segmentation accuracy due to uneven vertebral bone density, complex vertebral structure and low imaging resolution. To tackle these problems, this study proposes a segmentation method for spine CT images with a convolutional-deconvolutional neural network. The multi-scale residual module and the attention mechanism are introduced to improve the U-Net network, and the feature model is trained and tested. Experimental results on real data sets show that this method can effectively improve the accuracy and the efficiency of spine CT image segmentation. The estimated results of Dice coefficient and Intersection Over Union (IOU) are 0.97 and 0.94, respectively.

Abstract:Searchable encryption scheme supporting dynamic update has become a research hotspot to meet the needs of users for dynamic update of cloud documents. However, most of the known schemes update the index structure by the direct insertion at the tail, which leads to the leakage of the relationship between the newly added keywords and the document. Therefore, this study proposes a dynamic searchable encryption scheme based on semantic grouping. Firstly, the balanced binary tree is constructed as the index structure, and the number of nodes is reduced by semantic grouping for search efficiency improvement. Then, in light of the idea of partition matrix, virtual keywords are added to the matrix to ensure the security of update. Finally, the security of the scheme is analyzed by formal proof.

Abstract:Human pose estimation plays an important role in many computer vision tasks. However, it remains challenging due to complex pose changes, illumination, occlusion, and low resolution. The high-level semantic information from deep convolutional neural networks provides an effective way to improve the accuracy of human pose estimation. In this study, an improved stacked hourglass network is proposed. A large-receptive-field residual module and a preprocessing module are designed to better outline structural features of a human body so that rich contextual information can be obtained. The network performs well in the presence of partial occlusion, large pose change, complex background, etc. In addition, the positioning accuracy is further enhanced by the fusion of results from different stages. Experiments on MPII data sets and LSP data sets prove the effectiveness of this model.

Abstract:Oriented to object detection in optical remote sensing images, this study proposes an improved Single Shot multibox Detector (SSD) model aiming at typical objects, i.e., aircraft and car, in the images. First, a multi-scale feature fusion module is introduced to the SSD network model to fuse deep features and shallow features. As a result, more contextual information of features can be obtained and the network’s ability to extract object features is enhanced. Then, cluster analysis is performed according to the size distribution characteristics of target samples in the data set to obtain more accurate default bounding box parameters, thereby effectively improving the network’s ability to extract target location information. Finally, the proposed model is compared with SSD and YOLOv3 models on data sets common for object detection in remote sensing images, which demonstrates that the mean Average Precision (mAP) of object detection has been greatly improved and verifies the effectiveness of our model.

Abstract:The standard encoder-decoder model has a weak ability to extract time series data. For this reason, this study proposes a novel encoder-decoder model integrating the Bi-directional Long Short-Term Memory (Bi-LSTM) and Attention Mechanism (AM). First, the input data is extracted by Bi-LSTM from both positive and negative directions. Then, different weights are assigned to the obtained features by AM according to different moments, and corresponding background variables are generated according to different moments in the decoding stage. As a result, the airport passenger flow can be predicted. Further, the Shanghai Hongqiao International Airport is taken as an example for the experimental simulation with this algorithm. Experimental results show that compared with Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM), the proposed method reduces the average standard error by more than 57.9%. This study provides a new way for airport passenger flow forecast.

Abstract:Code management is an important part of software development. With the increasing complexity of software requirements and functions, multi-branch development scenarios become more common, which makes the difficulty of code management soar. Also, code leakage between branches arises, seriously affecting the development efficiency and version delivery quality. According to personal code management experience, this study analyzes commonly used branch management modes and explores the solution to the multi-branch code leakage mainly from perspectives of branch development and branch release mode. The practical application results show that the proposed solution can avoid the branch code leakage, optimize the code management process, and improve the overall code management efficiency.

Abstract:The loss of coding information, the poor adaptability to multi-scale building targets, and the insufficient contextual feature connection can be found in the classic Unet algorithm during the extraction of building features from remote sensing images. To tackle these problems, this study proposes a deformed-residual-pyramid codec network with multi-scale fusion. First, the original coding structure is replaced by the deep coding network and the down-sampling bypass network, which jointly extract the high-level feature information of the building target. Second, the residual pyramid structure combined with deformed convolution is introduced at the penultimate node of the coding network to improve the network’s ability to recognize multi-scale features and edge fuzzy features of buildings. Finally, the high- and low-level features are cascaded and merged layer by layer, and the segmentation result of the building is obtained at the end of the decoding network. The experimental results show that compared with the original model, the improved model has increased F1-score and MIoU by 1.6% and 2.1%, respectively.

Abstract:WeChat is one of the main applications of the modern Internet, but there are few studies on the characteristics analysis and modeling of its network traffic. The study takes WeChat traffic as the research object and finds that it has self-similarity and burstiness. In view of these two characteristics, we use the linear fractional stable noise model to characterize WeChat traffic and carry out the parameter estimation and the effect analysis of the model. The research results provide a basis for subsequent network performance analysis and traffic monitoring.

Abstract:The application fields of unbalanced data sets are becoming increasingly extensive, and the demand for them is getting higher. Taking the spectral clustering undersampling as a prerequisite, this study develops an unbalanced data mining method based on a self-encoding network to improve the classification accuracy of the overall data set. The clustering problem is converted into the multi-path partition problem of an undirected graph, and the spectral clustering is completed depending on the undirected graph and standardized processing. The majority of data sets are processed through selective undersampling to yield the classification boundary offset. The learning process is a self-encoding network of unsupervised learning, based on which the dimensionality of data is increased or reduced so that hidden features of each dimension can be obtained and the efficient representation and learning of data are realized at all levels. The self-encoding network is adjusted according to the comparison between the maximum mean difference and the preset threshold. The unbalanced data mining is then completed with the obtained classification interface. UCI data sets with different practical application backgrounds are selected, from which 10 sets of data are extracted as test sets. After spectral clustering undersampling, the simulation experiments demonstrate that the proposed method greatly improves the classification accuracy of the minority and overall mining performance, which shows good applicability and feasibility.