Abstract:Traditionally, Message Passing Interface (MPI) runtimes have been designed for clusters with a large number of nodes. However, with the advent of MPI+CUDA applications and GPU clusters with a relatively smaller number of nodes, efficient communication schemes need to be designed for such systems. This coupled with new application workloads brought forward by Deep Learning (DL) frameworks like Caffe and Microsoft Cognitive Toolkit (CNTK) pose additional design constraints due to very large message communication of GPU buffers during the training phase. In this context, special-purpose libraries like NVIDIA NCCL have emerged to deal with DL workloads. In this study, we address these new challenges for MPI runtimes and propose two new designs to deal with them: (1) a Pipelined Chain (PC) design for MPI_Bcast that provides efficient intra- and inter-node communication of GPU buffers, and (2) a Topology-Aware PC (TA-PC) design for systems with multiple GPUs to fully exploit all the available PCIe links available within a multi-GPU node. To highlight the benefits of proposed designs, we present the performance evaluation on three GPU clusters with diverse characteristics: a dense multi-GPU system RX1, with a single K80 GPU card per node RX2, with a single P100 GPU per node RX3. The proposed designs offer up to 14×and 16.6×better performance than MPI+NCCL1 based solutions for intra- and inter-node broadcast latency. we have enhanced the performance results by adding comparisons for the proposed MPI_Bcast designs as well as ncclBroadcast (NCCL2) design. We report up to 10×better performance for small and medium message sizes and comparable performance for large message sizes. We also observed that the TA-PC design is up to 50% better than the PC design for MPI_Bcast to 64 GPUs. The results clearly highlight the strength of the proposed solution both in terms of portability as well as performance.

Abstract:Recently, the training time of deep neural network has been greatly shortened because of the rapid development of computer technology especially the improvement of hardware conditions. The deep residual network has rapidly become a new research hotspot. The architecture exhibits good features in terms of precision and convergence. Researchers have delved into its nature and proposed many improvements on deep residual networks, such as wide residual networks, deep pyramidal residual networks, densely residual networks, attention residual network, etc. This study analyzes the construction of different residual units from the design of residual network, and introduces different variants of deep residual network. From different aspects, We compare the differences between different networks and the performance of these network architectures on common image classification datasets. Finally, we summarize these networks and discuss some research directions of future deep residual networks in the field of image classification.

Abstract:Corner detection is a basic subject in the field of machine vision and computer vision. Corner detection is sometimes called interest point. It not only simplifies the image information data, but also retains the more important feature information of the image to a certain extent. Corner detection includes three-dimensional scene reconstruction, motion estimation, visual tracking, image registration, and image matching. Computer vision has been widely used in the field of computer vision. This study classifies and describes the existing corner detection methods, which are mainly divided into gray-level intensity-based methods and edge contour-based methods. The other types of corner detection methods are also summarized, providing references for image corner detection technology.

Abstract:Improving the analyzing and understanding ability of the customer service system for group customers' electricity consumption problems seems to be one of the important ways to improve the quality of customer service for power industry. Based on clustering technology in data mining, this study establishes a customer service data analysis clustering model for customers' electricity consumption problems recorded by a customer service center, and then proposes an improved adaptive feature weighted K-Means clustering algorithm for the analysis of electricity consumption problems. The experimental results show that the proposed method can quickly and accurately realize the automatic clustering of customer service data and mine the hidden critical information of customers' electricity consumption problems, thus providing technical support for improving the quality of customer service and predicting the potential risk of customer service.

Abstract:Remote sensing image change detection is one of the hotspots of remote sensing application research. It has been widely used in urban change, environmental monitoring, land use, and basic geographic database update. Change detection is the feature and process of quantitative analysis and determination of surface changes from remote sensing data in different periods. The specific work is to analyze two or more images of different phases in the same region, and to detect the changed parts and unchanged parts. In this study, a change detection method based on stack noise reduction automatic encoder network is proposed. The deep learning algorithm applied to SAR (Synthetic Aperture Radar) satellite image change detection is improved, which is suitable for high-resolution remote sensing satellite image, and then improved on the structure of twin network. A change detection method based on branch convolutional neural network is proposed. Finally, the design algorithm removes the false changes such as shadow interference and noise, and tests it on the actual production data image of the high-resolution satellite 2 (GF-2). It has achieved sound results.

Abstract:In complex natural scenes, object recognition encounters the problems such as background interference, occlusion of surrounding objects, and illumination changes. At the same time, most of the identified objects have different sizes and types. In view of the above-mentioned problem of object recognition, this study proposes a medium or large size object recognition method based on improved YOLOv3 in unrestricted natural scenes (CDSP-YOLO). This method uses CLAHE image enhancement preprocessing method to eliminate the influence of illumination changes on object recognition in natural scenes, and uses stochastic spatial sampling pooling (S3Pool) as the downsampling method of feature extraction network to preserve the spatial information of feature map to solve the background interference problem in complex environment, and improves multi-scale recognition to solve the problem that YOLOv3 is not effective for medium or large size object recognition. The experimental results show that the proposed method has an accuracy rate of 97% and a recall rate of 80% on the mobile communication tower test set. Compared with YOLOv3, the algorithm has better performance and application prospects in object recognition applications in unrestricted natural scenes.

Abstract:In the process of detecting two-dimensional medical images, it is helpful for doctors to analyze image data more comprehensively and make accurate measures to deal with the illness by using the three-dimensional visualization technology of medical images. Taking the liver image as an example, firstly, continuous cross-sectional images in visible human dataset are extracted. An image segmentation algorithm based on region growing is used to extract the liver contour from continuous cross-sectional images. Secondly, interlayer interpolation of liver contour is used to construct a three-dimensional model of liver by using the VTK combined surface rendering method. Then the initial model is meshed to some extent to reduce data redundancy and complete the three-dimensional reconstruction of virtual liver. This work studies the liver solid texture synthesis with mapping based on the CUDA system, it also considers the reality for solid texture and balances the practical for virtual surgical instruments which plays a major role in the virtual liver surgery development.

Abstract:In view of the frequent external force damage incidents of transmission lines, manual inspection and traditional monitoring equipment cannot find the hidden dangers in time and effectively. Therefore, an intelligent monitoring system for transmission lines based on image recognition technology is proposed. The system uses convolution neural network depth learning algorithm to train the model, which can intelligently identify the potential safety hazards of transmission lines. A new intelligent monitoring mode is established, which includes front-end image acquisition, wireless data transmission, background recognition and analysis, and hidden danger directional push. In Foshan area, the system realizes 24-hour real-time monitoring and early warning of transmission lines, improves the monitoring ability of hidden dangers caused by external forces, and effectively prevents line tripping accidents caused by large-scale construction machinery.

Abstract:Balance wheel is commonly used in precision instruments. Its flatness plays an important role in the precision and accuracy of the whole instrument. Traditionally inspection of balance wheel is done manually. This study raises a flatness inspection system for balance wheel based on machine vision. First, the image of balance wheel is calibrated. Second, its sub-pixel edge information is extracted and boundary information is accurately calculated through projection mapping. The system supports continuous sampling and calculation and takes the dynamic range of the edge of the balance wheel as acceptability criteria of parts. The result of the test shows that this method has a satisfied real-time property and high accuracy, so it can meet industrial inspection needs.

Abstract:The logistics support of large units can not be separated from the process management information system at any time. The traditional methods adopt the fixed flow diagrams and the fixed forms, which can not meet the requirements of flexible and independent logistics management under the new situation. At the same time, the relational database mode adopted by traditional information systems is also difficult to adapt to the current big data in logistics support. This study designs and implements a logistics free process information system using Activiti and MongoDB. With the goal of free and flexible process management and the idea of relational database assisting non-relational database as design concept, the free process framework, key interfaces and important code logic are given, as well as the design method of two kinds of database collaboration. The implemented system effectively solves the above problems, greatly reduces the workload of developers, and provides an effective system design reference for logistics support process management informatization of large enterprises or units.

Abstract:In view of the current domestic panoramic roaming technology compatibility and poor implementation efficiency, this study proposes a panoramic technology based on WebGL, which eliminates the installation of rendering plug-ins. The system is more lightweight and has lower computer resource usage. Firstly, data is collected by combining the aerial photography by unmanned aerial vehicle and ground point shooting. Then, image splicing process is performed with scale-invariant feature transform. Finally, WebGL technology is adopted to combine with specific HTML5 framework to add scenes and make hot data, which are uploaded to the server to complete the publication. It is successfully applying to Guiping smart tourism construction project. Through this system, users can enjoy the scenery of scenic spots on the online virtual tour, and experience the immersion of panoramic roaming.

Abstract:Nowadays, the barrier-free device in the human-computer interaction system is generally extremely expensive and its volume is large, which fail to really walk into the daily lives of people with disabilities. To develop a set of low cost, high precision, portable barrier-free character input system, this study designs a set of character input system based on EOG (Electro-OculoGram), puts forward a flashing character interaction based on regional optimization, and optimizes the dynamic threshold algorithm for identifying target character algorithm. Experiments showed that the system has realized the identification accuracy of 97.73%, nearly 100%, using equipment which costs less than RMB1000, and average character input rate of 1.95 characters/min.

Abstract:In order to satisfy the demands of meteorological applications in prefecture-level city, aiming at the existing problems such as low reliability, poor scalability, and low concurrent processing ability of information system, this study adopts virtualization technology, load balancing strategy configuration, and linkage technology with virtualization. It designs and builds a virtualized resource platform to realize the expansion and recovery of dynamic resources. At the same time, this study designs and implements an application-based load balancing method and applies it to the actual meteorological domain. At present, the system has been put into operation. It provides stable and efficient application support for city and county meteorological applications. It obtains better application results.

Abstract:In this study, we use sensors, ZigBee wireless sensor network, embedded system, GPRS, cloud, big data mining analysis, and other technologies to design and implement a multi-indicator, multi-node, low-power “waters skynet monitoring system”. The test shows that the system design scheme is reasonable and feasible, and the system runs stably. It has realized distributed monitoring of water quality status, early warning, big data analysis and visualization of various scientific charts, effectively expands the scope of monitoring, accelerates the early warning and response speed, and improves the process of water quality monitoring and treatment.

Abstract:Based on Liaoning meteorological broadband wide area network, the HD video conference system of Liaoning Meteorological Department was established. The system follows the principle of advanced, standardized, safe and reliable, flexible expansion and easy operation, takes the provincial control center as the node, and has the function of high-definition video conference listening, watching, forwarding, and recording, which has been widely used in Liaoning meteorological service system. In this study, the Liaoning meteorological high-definition video conferencing system architecture, system composition, and key technologies to realize the system audio function are analyzed. At the same time, combined with the needs of business development, the new access system of mobile terminal video conference was set up. Finally, taking the audio fault of Liaoning Province HD video conference system as an example, the audio fault tree is established, and the audio stability of the system is quantitatively and qualitatively analyzed by using the fault tree analysis method.

Abstract:In the modern science and technology society, with the rapid development of digital image processing technology, image segmentation, and object edge detection are widely used in the medical, military, public defense, computer vision, and agricultural meteorology field. In this study, based on the classical Chan-Vese (CV) model, a piecewise constant image edge detection model with L1 norm data fitting term and TV2 second-order regular term is introduced. The new model uses a high-order regular function to penalize the objective function as a constraint on the new objective function, so that the model enabling to segment and detect images with low contrast and containing additional noise. Theoretically, we give reasonable assumptions, and a partial convergence analysis of the model is carried out. In terms of computation load, we study the theoretical solvability of the new model. Compulationally, for the numerical implementation of the model, the model is numerically solved by ADMM algorithm, and a new solution method is designed. A large number of numerical experiments were carried out with grayscale images and real images, and compared with the original CV model. The experimental results show that many advantages of the model with wide applications.

Abstract:With the increasing number of netizens, the users on the Internet has doubled, and a variety of comment data can be seen everywhere. So, it is very necessary to construct an efficient emotional classification model. This study combined Word2Vec with LSTM neural network to construct a three-class emotional classification model. Firstly, Word2Vec word vector model is used to train the emotion dictionary. Then, we construct word vectors for the current training set data by using emotional dictionary. Then, this study used the main parameters that affecting the accuracy of LSTM neural network model to train the model. The experiment found that when the data are not normalized, using the weight of He is initialized, the learning rate is 0.001, the loss function is mean square error, the RMSProp optimizer is used, the training rounds are 30, and the accuracy of traditional Word2Vec + SVM method improves by about 10%. The effect of affective classification promotes obviously, which provides a new way of thinking for LSTM model's sentiment classification.

Abstract:In the process of processing related network pictures, the traditional system is difficult to extract features and it makes the pictures recognition rate becoming inefficiency. In this study, we propose a module based on spatial transformation and dense neural network method to recognize the images, extract text feature, and transform the parameter about sensitive text pictures. The module using the deep GRU and CTC to mark characteristics of sequence prediction information, and serialization of dealing with the text can better improve the ability of wider text and fuzzy text information. Experimental results show that the recognition accuracy of the model in Caffe-OCR Chinese composite data set and CTW data set is 87.0% and 90.3% respectively, and the average recognition time reaches 26.3 ms/graph.

Abstract:Bridge crack detection has great significance for bridge condition monitoring. Distributed fiber optic sensors are widely used to detect bridge cracks. The state quo approach is based on Brillouin time domain analysis. Though being capable of measuring strain data across the surface of the structure, it has its own flaw—the strain anomaly at the crack damage is “submerged” and “confused” by the noise due to the lower signal-to-noise ratio of the measured strain data. To this end, we propose a classification detection method based on one-dimensional stacked convolution autoencoder, which are endowed with strong noise robustness and high resolution of auto extraction features. The proposed method consists of three steps. First, structural surface strain data is acquired by arranging fiber optic sensors, the fiber strain data preprocessed and the strain subsequence divided. Second, the characteristics of the strain subsequence are automatically extracted using a one-dimensional stacked convolution autoencoder. Finally, the extracted strain subsequence features are classified by a Softmax classifier into two categories—cracks or non-cracks. The method can effectively detect micro cracks and has high detection accuracy. Moreover, by experimental contrast we claim that the feature discriminability extracted by this method is better than that of convolutional neural network and stacking autoencoder.

Abstract:In order to improve the performance of the DNS server, a mathematical model based on M/M/c queue theory was proposed. The probability distribution function of response time was analyzed according to this model, which identifies domain name compression rate as the performance bottleneck. Due to the rule of traditional domain name compression algorithm, DNS servers can only perform real-time domain name compression when the query is answered, which causes a performance problem in the high-traffic scenario. To improve the domain name compression rate, the principle of domain name compression was analyzed from the aspects of DNS data characteristics. Based on this, combing with base relocation technology, a new domain name compression algorithm was proposed. The new design changes the traditional DNS data process, which reduces the real-time consumption during response by pre-compressing. Experimental results show that the algorithm improves the system resource utilization under the condition of small compression loss and achieves the goal of optimizing the percentile response time.

Abstract:Aiming at the shortcomings of current image restoration algorithms, such as discontinuity of repair effect, limitation of missing size, and instability of training process, an image restoration method based on generation antagonistic network is proposed. Using convolutional neural network, we can actually repair images of any resolution. In order to realize the real restoration effect of high resolution and the full learning of image features, we propose to obtain high resolution images based on the details and structure of DenseNet propagating source images, so as to realize the generation of missing images. As Iizuka et al. proposed the large computation amount generated by the expanded convolution part in the two-discriminator method, we propose to use JPU (Joint Pyramid Upsampling) to accelerate the calculation. Experiments in CelebA and ImageNet show that the proposed method can truly repair most broken images.

Abstract:Gait is a biological feature that can recognize identity at a long distance and without invasion. However, the performance of gait recognition can be adversely affected by many factors such as view angle, walking environment, occlusion, and clothing, among others. For cross-view gait recognition, the existing cross-view methods focus on transforming gait templates to a specific view angle, which may accumulate the transformation error in a large variation of view angles. To extract invariant gait features, we propose a method which is based on generative adversarial networks. In the proposed method, a gait template could be transformed to any view angle and normal walking state by training only one model. At the same time, the method maintain effective identity information to the most extent and improving the accuracy of gait recognition. Experiments on CASIA-B and OUMVLP datasets indicate that compared with several published approaches, the proposed method achieves competitive performance and is more robust and interpretable to cross-view gait recognition.

Abstract:K-means algorithm is a commonly used clustering algorithm and has been applied to traffic hotspot extraction. However, due to the number of clusters and the subjective setting of the initial clustering center, the traffic hotspots extracted by the existing clustering methods are often difficult to meet the requirements. Based on mutual information and divergence, an improved SK-means algorithm is proposed and applied to traffic hotspot extraction. In the proposed method, an initial clustering center is found based on mutual information between different points. In addition, the number of clusters is determined based on the ratio of mutual information and divergence. The proposed method is applied to the extraction of traffic hotspots in Chengdu for a certain period of time, and compared with the traditional K-means, the experimental results show that the proposed method has higher clustering accuracy and the extracted hotspots are more realistic.

Abstract:The intelligent traffic application under the single camera scene has been well developed, but the cross-regional research is still in its infancy. This study proposes a cross-camera scene stitching method based on camera calibration. Firstly, the mapping relationship between the physical information in the sub-world coordinate system and the two-dimensional image in the two camera scenes is established by vanishing point calibration. Secondly, the projection transformation between the cameras is completed by the common feature information between the two sub-world coordinate systems. The road scene stitching is completed by the proposed inverse projection idea and the translation vector relationship. The experimental results show that the proposed method can achieve road scene splicing and cross-region road physical measurement, which lays a foundation for relevant practical application research.

Abstract:At present, the interference in the moving target tracking task is very deceptive, and the target tracking algorithm is easily deceived by the data set with traps. In order to improve the tracking algorithm's effect on tracking dataset, this study proposes an improved DPP-SiamFC tracking algorithm based on SiamFC twinning network. This algorithm introduces DPP (Detail-Perserving Pooling) pooling layer and residual network based on the original network, effectively retaining the details of the target. This study also verifies network performance on the VOT2017 tracking dataset, the experimental results have achieved the goal of improving network performance.

Abstract:Vehicle logo location is one of the key technologies of vehicle logo recognition system. However, because of the different texture and variety of radiators in the background, it is difficult to locate the vehicle logo. Therefore, a vehicle logo location method based on background texture is proposed. Firstly, the method locates the vehicle logo roughly according to prior knowledge, then divides the background of the vehicle logo into three categories according to its characteristics on horizontal and vertical projections, and then uses Sobel operator to ablate the background of different types of radiators. In order to better remove the influence of radiator background on the location of the vehicle logo, a neighborhood binarization method is introduced, which combines projection-based method. The denoising method further deals with the noise points, so as to realize the accurate positioning of the vehicle logo. This method is suitable for different types of vehicle logo background conditions. The experiment results show that the proposed algorithm has higher accuracy and applicability by positioning 1000 images, and the overall positioning accuracy can reach 97.10%.

Abstract:With the continuous development of smart grid, the automatic reading system of electric energy meter based on digital image processing method is widely used. To improve the accuracy of automatic reading of traditional electric energy meter, a new method of automatic reading of electric energy meter based on YOLOv3 network is proposed. For the electric energy meter image, a counter positioning model based on the YOLOv3-Tiny network is constructed and trained, the trained target model is used to locate the counter target area, and the counter area is generated to achieve a counter image. For the counter image, a counter recognition model based on the YOLOv3 network is constructed and trained, and the trained model is used to identify the number of the counter target area. The electric energy meter data set published by the Federal University of Paraná Brazil was selected as the research object. The comparison experiment with YOLOv2-Tiny positioning model and CR-NET recognition model shows that the proposed method has higher positioning accuracy and recognition accuracy.

Abstract:An important application scenario for target detection is the detection and location of indoor mobile personnel. In this study, we propose an indoor personnel detection method to improve YOLOv3. First, the proposed method clusters the dataset by using K-means algorithm and designs a DE-YOLO deep convolutional neural network structure. Through the dense connection in the DE-YOLO network structure, the compression of the model sizes and the reuse of the feature information are realized. Finally, the target detection is performed on the extracted features. Experiments show that the application of the newly improved deep convolutional network has greatly improved application effect on VOC2012 datasets.

Abstract:Specular highlight removal is a hot topic in the field of computer vision. Existing methods based on dichromatic reflection model,which separate diffuse and specular reflection components to remove specular highlights in a single image, tend to cause color distortion and texture loss. To relieve this problem, the pixel clustering algorithm is improved by using pixel intensity ratio to remove specular highlight, which can more accurately classify pixels and improve the image color distortion. Firstly, the difference between the original image and the single channel image of the minimum intensity value is calculated, to obtain the specular-free image. Secondly, the minimum and maximum diffuse chromaticity values for each pixel in the highlight area is estimated based on the specular-free image. Finally, the distribution pattern of the pixels in the highlight area are analyzed in a minimum-maximum chromaticity space and clustered by x-means method. The specular components of highlight area pixels can be easily separated by using the estimated intensity ratio of diffuse pixels, thereby getting a non-highlight image. Experimental results show that, compared with the existing method, the peak signal-to-noise ratio increases by 2% to 4% on average, and the image color distortion and texture loss are improved with better visual effects.

Abstract:In some traditional source location algorithms, the requirement of the initial source solution is high, the dependence is large, the search range has a certain limitation, and the source location optimization is difficult to be carried out in the large area range where the group wave aliasing is serious and the spectral component is complex. In order to solve this problem, a method of underground shallow source location based on Quantum Particle Swarm Optimization (QPSO) algorithm is proposed. The algorithm is realized by simulation, and the advantages and disadvantages of the proposed algorithm and the traditional algorithm are evaluated. The experimental results show that in the range of -100 m×100 m×-40 m, the location algorithm's positioning accuracy of the shallow single-target source based on QPSO is obviously higher than that of the traditional source location algorithm based on PSO, the positioning accuracy can reach 0.324 m, which is of great practical application value.

Abstract:The network community hierarchies are defined by heterogeneous of different scales of link density in essence, it is necessary for network community to detect the dynamically changing information during hierarchies division. In view of this, a method of extraction of network community hierarchies based on spectrum top-segmentation is proposed. Firstly, the spectrum top-segmentation is defined as a dichotomy of subnetwork that no any top-level community can cross two parts, and an expected division top-level segmentation is presented. Then, the queue and link-density are introduced to decompose network, and an algorithm of network community levels extraction is presented. The simulation result shows that the performance of proposed method is better than that of synchronization and multi-scale in stochastic hierarchical networks, and the method is applicated in Email real-world network effectively.

Abstract:Verifiable Secret Sharing (VSS) is an important cryptographic primitive in distributed computing. The most of pervious VSS almost depended on the commitments which was established under the computational assumption of discrete algorithm problem which had been proofed unsecure. So a quantum-resistant VSS scheme which can be applied in secret sharing schemes implemented by different methods is called for. In this study, we analyze the existing verifiable sharing schemes. In order to solve the flaws in the existing schemes, we propose a new scheme with the applicability to secret sharing implemented by lattice cryptography. In addition, it has higher verification efficiency compared to past schemes and resistance so far to cryptanalysis by quantum algorithms.

Abstract:With the continuous development of digital resources, facing large-capacity digital resource verification, the existing file verification methods have greatly improved compared with the traditional file verification in terms of execution efficiency and utilization of server resources, but there are still some characteristics such as inflexibility, waste of server resources, limitations of file verification, etc. Therefore, this study proposes a new method based on server resource flexibility and concurrency. By using multi-concurrent and block-reading methods, we can rationally utilize the resource utilization of servers to improve the efficiency of file verification under the condition that the theoretical complexity of file Hash calculation remains unchanged.

Abstract:Echo State Network (ESN) owns simple network structure and is coupled with a time parameter and thus it shows important theoretical and application values in time series forecasting. In this study, we propose to optimize the output weight matrix by Adaptive Backtracking Search optimization Algorithm (ABSA) to overcome overfitting problem caused by linear regression algorithm. ABSA adopts adaptive mutation factor strategy to replace the strategy of randomly given mutation factor in standard BSA to achieve the balance between convergence accuracy and convergence rate. Experimental results show that the ESN optimized by ABSA outperforms the basic ESN without optimization and the ESNs optimized by other EAs.

Abstract:In order to improve the forecasting software aging trend, a New Particle Swarm Optimization Simulated Annealing algorithm (NPSOSA) is proposed to optimize the weights and thresholds of BP neural network, and then NPSOSA-BP neural network forecasting model is constructed. The software aging test platform was built to collect the required aging data and conduct simulation training. The experimental results show that the NPSOSA-BP neural network model improves the prediction accuracy and applicability compared with the BP neural network model optimized by the traditional Particle Swarm Optimization (PSO) and the traditional Particle Swarm Optimization Simulated Annealing algorithm (PSOSA). The validity of this method is verified in this application field.

Abstract:The channel environment of high-voltage transmission line has a great impact on the safety of high-voltage lines. In the past, manual inspection of the channel environment was a necessary way. Nevertheless, the manual inspection is dangerous and difficult, and its efficiency is low. To solve the problem, we propose a super-pixel combined with deep neural network for high-voltage transmission line environment detection. First, we obtain the overall image of the channel environment by the splicing technique for UAV aerial photography. Then, we employ the super-pixel segmentation algorithm to preprocess the image, in which we choose the SURF descriptor to extract the superpixel features because of its rapidity and effectiveness. Finally, the deep neural network model is used for training and classification and the superpixels are classified to achieve the purpose of detection. The experimental results on real environment images show that the inspection efficiency of high-voltage transmission line channel environment is greatly improved and the proposed algorithm is effective.

Abstract:In this study, we propose a one-sample person re-identification method, which adopts a progressive learning framework in the process of iteration in order to making full use of the characteristics of labeled data and unlabeled data to optimize the model. In this framework, we iteratively train convolutional neural network to update the model and utilize multiple-model training together to select the reliable pseudo-label data during label estimation. Then, we update training data for the next round of training. The training data is splited into three parts: labeled data, pseudo-labeled data, and indexed-labeled data. We set up the corresponding loss function for each set of data and update the CNN model by the joint training on the three parts. In the progress of iteration, the pseudo-labeled data and index-labeled data are constantly updated. Under the one-sample set, rank-1=65.3, mAP=45.6. When the rate of labeled data is increased to 40%, rank-1=83.8, mAP=64.9. The result indicates that the semi-supervised person re-identification method proposed in this study can provide excellent results comparable to the supervised learning method with less labeled data, which fully demonstrates the effectiveness of the method.

Abstract:Incomplete information game is an important research direction in the field of artificial intelligence. In this study, we introduce a solution of incomplete information game based on Satisfiability Modulo Theory (SMT). We describe the dynamic process of the game as corresponding constraints through Situation Calculus and write the constraints into propositional logic formula. Then, we convert the reasoning problem to the problem of satisfiability of logical formula and call the SMT solver. The application shows that the proposed algorithm can effectively deduce the answer of the game.

Abstract:This work analyzes the difficulties and limitations of medical plant training, investigates and studies the suggestions of the teachers and students of traditional Chinese medicine colleges and universities and Chinese medicine lovers on the practical training, combining with virtual reality technology, using Unity3D simulation to reproduce the real medical plant training scene, to realize the intertemporal and temporal learning of the function and taste of traditional Chinese medicine, to understand the characteristics of traditional Chinese medicine, to understand the growth environment of medicinal materials, to distinguish the characteristics of medicinal materials and to identify their characteristics, and to collect the whole strains of traditional Chinese medicine or the sites of medicine. In order to enhance the interest of learning, four-season variation, diurnal alternation, and different weather are added to optimize the environment. In addition, different environments can publish tasks and complete tasks in the form of games for reward. Applying virtual reality to practical teaching training can improve learning efficiency, increase professional knowledge, cultivate self-rescue ability, effectively exercise the safety consciousness of users, and reduce the occurrence of safety accidents.

Abstract:A plan deduction system is developed to support multiple real-time data transmission satellite reconnaissance missions. The overall structure and information flow of system are designed, followed by research about the key technologies needed for system development. A data transmission task planning algorithm for area reconnaissance, a satellite area coverage analysis algorithm based on grid, a visualization method for target area and reconnaissance scope based on OSGEarth are proposed. The developed planning and deduction system has been successfully applied to an unnamed task to carry out the task planning, capability analysis, and process deduction.