Abstract:Autoencoder, as an important branch of deep learning, has appealed many outstanding researchers in this field. Researchers studied its essence, and proposed many optimized approaches, such as sparse autoencoder, denoising autoencoder, contractive autoencoder, and convolutional autoencoder. After reading a number of articles on autoencoder methods, we found that the optimized autoencoder had sound experimental results in terms of image classification, natural language processing, and object recognition. Therefore, this review analyzes the basic principle and structure of optimized autoencoder in details. In addition, the multi-perspectives evaluation and analysis on the experimental results in literatures are carried out as well.

Abstract:Traditional Data Center Network (DCN) cannot satisfy large-scale networks and cloud-oriented infrastructure. Routing policies in traditional DCN complicate the forwarding elements and the topology management is inflexible in large-scale DCN. Protocol Oblivious Forwarding (POF) is a groundbreaking forwarding plane technology. In this study, we proposed two technologies of DCN. We designed extremely simple routing mechanism with POF. Forwarding elements of DCN can process packets easily using simple source routing protocol. Furthermore, we designed a cooperative topology management strategy to reduce redundancy of probe packets. We implemented our technologies in POF-based DCN and conduct some experiments. Results demonstrate that our routing mechanism can effectively simplify the forwarding elements and our topology management strategy can reduce the probe packets redundancy.

Abstract:For the situation that the adjacent markers of Motion Capture (MOCAP) data missing for a period of time due to lights and other factors when practically gathering data, a new MOCAP data recovery algorithm is proposed by using the latent correlation and the skeleton constraint in MOCAP data. The algorithm firstly transforms the MOCAP data to represent the changes of the relative position of adjacent markers to acquire the skeleton constraint term. Then the sparse representation and the skeleton constraint term are used for dictionary training which is utilized to recovery missing data. The experiment results show that the algorithm can improve the recovery accuracy of the coordinates of the missing markers and increase the bone length recovery accuracy to 10^-4 cm, and verify the feasibility and effectiveness of the algorithm.

Abstract:In this study, we propose an Overlay security mechanism, which is secure and practical, based on Space Operating System (SpaceOS). Firstly, the system security domain is divided, and the requirement of security mechanism is defined. The design idea of security mechanism is put forward, and the security requirement is proved by formal method. Then, the feasibility of the technology is explained by Overlay file system, and the security mechanism is applied to SpaceOS. Finally, the performance test, security test, and practicability test are carried out. The test shows that the Overlay security mechanism is secure and practicable. It has little effect on the speed of the SpaceOS has practical value.

Abstract:These dominant algorithms to learn a similarity is the metric learning that learns a Mahalanobis Similarity Function (MSF) to estimate the similarity of a pair of persons. However, the MSF only projects a pair of persons into feature difference space and ignores the appearance of each individual. In this study, we proposed to learn a Bidirectional Relationship Similarity Function (BRSF) that greatly strengthens the modeling ability of the similarity function. BRSF not only represents the cross correlation relationship of a pair of persons, but also describes the auto correlation relationship. We use the ideal of the Keep It Simple Straightforward Metric (KISSME) algorithm to learn a similarity function. Specifically, the auto correlation relationship and cross correlation relationship of a pair of sample features are expressed by Gaussian distribution. Finally, by converting the ratio of the final Gaussian distribution into the form of BRSF, we get a similarity function which is robust to the change of background, viewpoint, and posture. The proposed method is demonstrated on two public benchmark datasets including VIPeR and QMUL GRID, and experimental results show that the proposed method achieves excellent re-identification rates compared with other similar algorithms. Moreover, the re-identification results on the VIPeR dataset with half of dataset sampled as training samples are quantitatively analyzed, and the performance of the proposed method achieves a 53.21% at Rank1 (represents the correct matched pair).

Abstract:A new algorithm for keyword recognition based on audio automatic segmentation and depth neural network is proposed to identify the requirements of keyword recognition on the condition of low or zero resource. Firstly, an improved speech segmentation algorithm based on metric distance is used to divide the continuous speech stream into isolated syllables, and then the syllable is subdivided into short audio segments which are connected with the phoneme state. The segmented audio segment has the characteristics of large difference between the segments, and the characteristic variance of the segment is small. Then, an improved vector quantization method is used to encode the state features of the audio fragments, and the high precision quantization coding and the low precision quantization coding of the words are realized. Finally, the syllable is used as the recognition unit, and the compressed state transition matrix is used as the whole feature of the syllable. It is sent into the deep neural network for speech recognition. The simulation results show that the algorithm can identify many specific keywords from the natural speech stream, and the algorithm is easy to understand, the training is simple and the robustness is better.

Abstract:Utilizing the existing methods of legacy-to-cloud migration, considering the special requirements on the stability, real-time, high concurrency, and security aspects of land and resources online trading system, this study proposed a solution to maintain the above characteristics after the migration. The solution is adapted to the cloud from data exchange between internal and external networks, transaction engine, data exchange mode of Browser/Server, and so on. Finally, the test and run are formally carried out in Chongqing, Jiangxi, and other places. This method can also be used as a reference for the legacy-to-cloud migration in public security, urban planning, and other departments.

Abstract:This paper presents a description of the intelligent-lock model which could remotely share its key through smart phones. It also provides a set of hardware models that make the whole system function normally. The facility of sharing the keys is surprisingly rare in today's field of intelligent locks. Instead, the combination of different means, such as passwords, access cards, fingerprints, and Bluetooth, is widely used to unlock intelligent locks. But as is shown in this paper, the RSA algorithm makes a better effect on realizing the keys' confidentiality and timeliness while sharing the keys. As for hardware system, for instance, used for unlocking through fingerprints or access cards, we can also find a workable design based on the controller made by low-power STM32F103 single chip. The result shows that the system can actualize the function of sharing keys well.

Abstract:The most important issue of agricultural informatization is realizing the agricultural data collection and processing. The node design and deployment of agriculture wireless sensor network and the PC software design form the network system in this study for improving the informatization of the fruit cultivation, thus realize the environmental information collection in farmland collection system, therefore reduce the cost of the traditional collection technology through the network, solve the problems of high cost, difficult layout, cumbersome equipment, and maintenance difficulties, and realize the remote information management.

Abstract:With the rapid development of the communication market and the high penetration rate of mobile phones, it is possible to use mobile phone data to study the development of human activities and urban planning. The main work of this study is to build a big data real-time processing analysis platform, and based on this platform proposes a method to understand the urban population distribution by using of mobile data. Experiments show that dynamic population perception based on mobile phone signaling data can reflect the actual urban population distribution, which is of great significance to urban traffic regulation and public resource allocation optimization.

Abstract:In view of the current situation of China's relatively low informatization level of biomass energy engineering, we combine today's popular big data processing technology, design and implement a biomass energy engineering data resource management platform based on Hadoop open source framework. This paper introduces the main function modules of the platform, including the management of biomass energy engineering, the management of monitoring indicators, real-time monitoring, the management of anti-fraud model, statistical analysis, etc. In-depth research on key technologies of platform construction has been carried out, including data acquisition, big data storage, big data processing, load balancing, and so on. Combining Internet of Things technology, Internet technology, and big data processing technology with biomass energy engineering organically, we can improve the informatization level of biomass energy engineering, ensure the production safety, optimize the manufacturing process, and maximize the benefits, thus provide theoretical and practical basis for the construction of similar projects.

Abstract:For the purpose of developing virtual CA6132 lathe teaching software, adopted the virtual simulation method, a lathe teaching system software with virtual reality features is developed. The general process of virtual lathe development is put forward, the facade segmentation technique for correcting images is described, the virtual Lathe Simulation System based on Unity3D engine software is developed, the basic principles of virtual reality are explained, the three-dimensional model of CA6132 lathe is established, the lathe cutting simulation method is put forward under Unity3D engine software, the virtual reality Lathe Simulation System based on virtual reality equipment zSpace has been developed, and a more natural way of scene interaction in virtual reality is proposed. The simulation results show that the virtual reality teaching system of lathe has sound teaching effect, and reduced risk of actual lathe operation.

Abstract:This paper introduces a replay style testing tool that could automatically test HXDSP related basic software kits. The testing tool mainly tests debugger host software, and can also test other related software such as compiler, simulator, and so on. It is convenient to add test cases and relay test cases by groups on the testing tool. The testing tool will significantly reduce workload of testing basic software kits on HXDSP.

Abstract:Turbo code and LDPC code are traditionally employed in the channel coding, which is used for the communication between Unmanned Aerial Vehicles (UAVs) and ground receiving equipment. Unfortunately, the disadvantage of both turbo code and LDPC code is high complexity of decoding, high time-consumption, and high-priced hardware cost. Moreover, LDPC code may lead to error floor at high Signal-to-Noise Ratio (SNR). Oppositely, it is more efficient for decoding using Gray code in UAVs. The reason is that it has lower computational complexity and shorter operation time. Furthermore, for hardware development, it is able to achieve simple and low-power architecture. Thus, in this study, a UAV image transmission decoding algorithm is proposed based on adaptive hard/soft decoding of Gray code. In this algorithm, the parity bit decoding mechanism is designed for switching between hard- and soft-decoding. The simulation results indicate that the algorithm is efficient and reliable. It is also proved that the proposed algorithm can be adaptive to different requirements of image precision through adaptively selecting the suitable decoding method, which also speeds up the procedure of decoding significantly.

Abstract:With the rapid development of information technology, the need for data storage is growing, and the requirements for I/O performance are getting higher and higher. Compared with the traditional Hard Disk Drive (HDD), Solid-State Disk (SSD) has more advantages such as high availability, low energy consumption, no seek time, etc. Due to these reasons, SSD gradually replaces HDD to become mainstream storage media. The OpenSSD project is an initiative to promote research and education on the recent SSD technology by providing easy access to OpenSSD platform on which open source SSD firmware can be developed. In this study, we investigated the Flash Translation Layer (FTL) algorithm on the Cosmos OpenSSD platform, analyzed all kinds of factors that affect the I/O performance, designed a way to carry out the buffer management and flash management, and finally improved the I/O performance effectively.

Abstract:To improve the speed and precision of registration in image stitching, this study proposes a modified RANSAC algorithm based on Determinantal Point Processes (DPP), aiming to tackle the issue of robustness model estimation. This method utilizes global negative correlation of the DPP sampling to model matching feature points, eliminates those incorrect matching points, and therefore realizes the homogenization and decentralization of the sampling. The point set extracted in DPP is used as the input of RANSAC to elicit transformation matrix. Experimental results show that compared with traditional RANSAC algorithm, this algorithm ensures higher accuracy and robustness, which greatly enhances the efficiency of automatic image stitching.

Abstract:In automatic mushroom picking process, some influencing factors such as the presence of various backgrounds in the image, the huge scale difference among different mushrooms especially in overlapping mushrooms, make it tough to locate mushroom. In order to improve the target location accuracy, a new method of complex background suppression based on Harris corner detection was put forward. In view of the scale differences among overlapping mushrooms, an iterative algorithm for seeking extreme points of distance map was proposed. These extreme points were used as seeds in the Watershed algorithm for the segmentation of overlapping mushrooms. Finally, the segmented image for each overlapping mushroom was processed with the method of elliptical fitting to get the contours and center coordinate of individual mushroom. In order to verify the algorithms proposed in this study, an experiment was conducted over the mushrooms grown in the lab. The test results reveal that the location detection success rate is 86.3%. The average image processing time is 0.711 s that is aligned with the requirement of the automatic mushroom picking system.

Abstract:The user-based collaborative filtering recommendation algorithm is based on the calculation of the similarity between users when the neighbor user is screened, and the increase in the amount of data exacerbates the sparseness of the data, which leads to the poor accuracy of the results and affects the recommendation accuracy. Aiming at this problem, this study proposes a recommendation algorithm based on the combined similarity of users. The calculation of combined similarity of users is divided into two parts:the similarity of the user's preference for item attributes and the similarity of the demographic information between the users. The algorithm introduces the LDA model to calculate the preference for the user's item attribute, and the scoring data is only used as the screening basis when calculating so as to avoid using it directly as well as slow down the influence of sparse data on similarity calculation results. While the similarity between demographic information is measured by Hamming distance after the numerlization of demographic information. Experimental results show that the proposed algorithm is superior to the traditional collaborative filtering recommendation algorithm in recommendation accuracy.

Abstract:The 0-1 knapsack problem is a classical NP complete problem, which has a wide range of applications in real life. In view of the existing algorithms with the shortcoming of low precision in solving 0-1 knapsack problem, an inducing factor monkey algorithm is proposed to solve the problem. The basic idea of the proposed inducing factor monkey algorithm is that an inducing factor is adopted in the climbing process of the basic monkey algorithm to induce it to crawl upward, which can escape from local optimal solution and find the global optimal solution. In the simulation, compared with the existing methods, the results show that the proposed inducing factor monkey algorithm for solving 0-1 knapsack problem is valid.

Abstract:The research of personalized learning model in big data environment is a hot research topic under large-scale network learning environment. In view of the shortcomings of the traditional intelligent test paper generating strategy, such as the lack of data training, personalized features are not prominent, and the uneven distribution of knowledge points, etc. This study puts forward a personalized practice model, optimizes the rules of paper organization, and makes the individual characteristics more accurate. To a certain extent, it helps student to understand and digest the weak points and blind spot. In this paper, in order to develop a personalized learning practice strategy for personal learning, the knowledge points of each chapter will be transformed into tree management, and add the knowledge point error rate element into the knowledge tree. Finally, this new research model is applied to teaching and education system for experimental research. Research shows that the improvement of this key point is more conducive to improve students' overall academic achievement in general.

Abstract:Compared with the dictionary learning and recognition method based on sparse constraints, the projective Dictionary Pair Learning (DPL) has faster learning speed and a higher recognition rate. In order to further improve the recognition ability of DPL, this study proposes an improved DPL algorithm K-DPL, which combines Kernel Principal Component Analysis (KPCA) and DPL method. In K-DPL, by using the kernel technique, the samples are mapped to high-dimensional space to solve the nonlinear problem, and then DPL is used to get more discriminant dictionary by training the dictionary. Experiments on ORL datasets show that recognition rate is increased by at least 1.5% and the recognition speed is increased by about 20 times compared to DPL at different training ratios by using K-DPL. On the extended YaleB and AR datasets, compared with DPL, the recognition rate is increased by 0.3% and 0.4% respectively, and the recognition speed is improved by using K-DPL as well. It indicates that K-DPL has good robustness to illumination and occlusion.

Abstract:In order to obtain the ideal wireless network information detection results, a wireless network channel mechanism based on Least Squares Support Vector Machines (LSSVM) is proposed. Firstly, the research on the current situation of wireless network channel detection is analyzed, and the hypothesis model of wireless network channel detection is established. Then, using the least squares support vector construction of wireless network channel detection model, the particle swarm algorithm of LSSVM parameters are optimized. Finally, the wireless network channel detection experiments on MATLAB 2014 platform are performed in order to verify the effectiveness of the wireless network channel detection. The results show that the LSSVM for the wireless network channel achieves high precision detection results, the wireless network data transmission success rate is improved, and the error rate of data transmission is greatly reduced. Under the same experimental conditions, the wireless network channel detection results are significantly higher than that of the current classical detection mechanism, which verifies the superiority of the proposed model.

Abstract:The plasma disruption always happened during EAST discharge experiment, disruption shot distinction and parameters extraction are significant in disruption physics and future ITER current quench time prediction. In order to view and analyze history disruption shots conveniently, we distinguish disruption shots and extract crucial parameters based on the value of plasma current using MATLAB software and archived them in MySQL database. Meanwhile, the websites in which contrast display preset parameters and actual discharge parameters were built by adopting Dygraphs. The implementation of EAST disruption distinction and waveform display will ultimately lead to reduce the triviality of manual record, facilitate query and analysis, and improve work efficiency of technicians.

Abstract:In this study, a non-contact measurement method with interactive function is proposed by using Microsoft MixedReality product HoloLens. Measurement space can be scanned by using HoloLens depth camera, the target points can be determined for non-contact measurement purpose through gazing and gestures, calculation of the distance, area, and volume between the target points can be performed automatically, in addition to interact with the speech recognition function. Compared with traditional measurement tools, interactive measurement tools developed in this study allows non-contact measurement by using more natural interactive mode, the measured range became more widely, not only for measuring the distances, but also for calculating the areas and volumes automatically. The places which are inconvenient reached by the traditional tape measurement can be measured by non-contact measurement method proposed in this study, and which has certain application value.

Abstract:In order to improve the recommendation efficiency of collaborative filtering algorithm, this study proposes a collaborative filtering parallelization algorithm based on Hadoop platform. The traditional user-based collaborative filtering is carried out under Hadoop platform for MapReduce Programming model, to achieve parallelization. By using the MovieLens common data set to improve the comparison before and after the algorithm, verify that the parallel collaborative filtering efficiency is higher, and also more suitable for large-scale data recommendation.

Abstract:In this study, a scene image classification algorithm is proposed which combines Gabor-LBP frequency domain texture features and lexical model semantic features. The frequency domain information which obtained by Gabor transform, the corresponding LBP feature, and semantic features which extracted by the visual Bag Of Words (BOW) package model are fused to realize the classification. In order to verify the algorithm, we use two standard image test datasets to compare and test. The experimental results show that the proposed algorithm has obvious advantages in improving image texture expression, especially for image illumination, rotation, and scale.

Abstract:Wavelet transform image compression coding method in the high compression ratio of the reconstructed image quality is often poor. To solve this issue, a coding algorithm based on spectral graph wavelet transform is proposed. In this method, the image is transformed into a graph, the spectral graph wavelet coefficients are obtained by using the spectral graph wavelet transform to decompose the graph, the energy of these coefficients is attenuated with the increase of the scale. Then, the SPECK algorithm is improved according to the characteristics of the spectral graph wavelet coefficients. Finally, the spectral graph wavelet coefficients are quantized, and the quantized coefficients are compressed by the improved SPECK algorithm, and the original image is restored from the sparse coefficients while the amount of image data is compressed. The experimental results show that the coding method is effective for natural image compression, compared with the compression method based on wavelet transform, the PSNR of the reconstructed image is improved and the change is smooth, and has a larger compression ratio at the same time.

Abstract:Aiming at the defects of the existing mirror, by using the hyperbolic mirror design, the mirror is divided into the main view area and side view area (the main view area closes to the plane mirror, and the side view area is convex). In order to meet the demand of smooth transition for two regional mirror, a surface fitting model is set up to fit the complex surface by triangle based linear fitting the algorithm, thus the mirror surface shape optimization is realized. The optimization is mainly manifested in perspectives of those the visible area becomes larger, the blind area decreases, and the image distortion rate is small. During the transition of the transition line, the image is coherent and the image distortion rate around the body is below 3%.

Abstract:The positioning of finger knuckle in hand image is the basis for the hand gesture recognition and motion tracking in intelligent assembly by human-computer interaction with machine vision. The accuracy of hand knuckle information has direct influence on the gesture description and behavior recognition. Considering the random distribution characteristics of knuckle images, the image preprocessing has made by homomorphic filtering to enhance the image features. Classification feature learning of clustering set on knuckle image is finished based on the Gaussian process model. The characteristics model of the data distribution is learned with the clustering measure of the sample. The two type feature model from the feature learning is a detector for image feature, and the detecting results are the two likelihood values of the image. The finger knuckle target is recognized directly according to the estimation results while the two types model likelihood value as the input value which is marked by the positive and negative samples. The hand knuckle recognizing experiments with different location are held, and the knuckle detection is carried out with our own created test knuckle library. It shows from the experimental results that the posterior probability distribution can be obtained directly, and the recognition accuracy and efficiency of target are improved. The algorithm presented above is feasible.

Abstract:A collaborative filtering recommendation algorithm based on news timeliness is proposed. Firstly, by analyzing the characteristics of the news timeliness, the timeliness model of news is established. Then, the user-based collaborative filtering algorithm is improved combining the news timeliness model. Finally, the experimental results show that this method can highly enhance the performance of user-based collaborative filtering algorithm, and ameliorate the accuracy and recall rate of news recommendation.

Abstract:Aiming at the weaknesses of sparse data, low scalability and large computing existing in the current collaborative filtering algorithm, a BlockClust-Alternating least squares with Weighted regularization (BC-AW) collaborative filtering recommendation algorithm is proposed. Firstly, the user and the item of the original scoring matrix are jointly clustered and several submatrixes with the same scoring mode are generated. According to the research, the scale of these submatrixes is far less than the original scoring matrix which effectively decreases the computational complexity in the prediction process. Then, the regularized iterative least-square method is applied to each submatrix to predict its score. Hence recommendation is realized. The simulation results reveal that the proposed algorithm can effectively improve sparsity, expand scalability, and reduce computing compared with the traditional one.

Abstract:To solve the urgent technical requirement of face recognition in security industry, this research and implementation of face recognition based on deep learning for identification photos are carried out. The key technologies and algorithms of face recognition are implemented and compared. At the same time, to further improve accuracy, the method of feature remapping using Siamese network is proposed. Finally, the experimental results show that by using deep learning algorithms and face-library constructed by identification photos, efficient and accurate face recognition is achieved.

Abstract:To address the problems such as the inherent sparsity in the short text and the "lexical gap" of traditional classification model, using Word2Vec model to map words to a spatial vector of low-dimensional real number according to context semantic relations can effectively ease the sparse feature issue of short text. However, further study found that only using Word2Vec will ignore the influence of different parts of speech on the short text. Therefore, we introduce part of speech to improve the feature weighting approach, in which the contribution of speech is embedded into the traditional TF-IDF algorithm to calculate the weight of the words in the short text, and the vector of short text is generated by combining the word vector of Word2Vec. Finally, we use the SVM to achieve short text classification. Experimental results on Fudan University Chinese text classification corpus validate the effectiveness of the proposed method.

Abstract:Integrated Modular Avionic (IMA) has become a popular aircraft avionics system, ARINC653 is a standard interface for IMA architecture of aircraft avionics equipment, the real-time scheduling for ARINC653 partition Operating System (OS) is a key issue. There are lots of scheduling algorithms for ARINC653 partition OS based on single-kernel processor. In this study, a scheduling method based on Multi-kernel Load Proportional Round Robin (MLPRR) is proposed. The method calculates the task weight according to the load ratio of the task, and completes the scheduling of the task on the multi-kernel processor to meet the real-time requirements of the multi-kernel partition OS. Experiment results show that MLPRR is feasible and efficient.

Abstract:To study the evacuation situation of underground miners in mine disaster, an escape model of mine fire was established by using Agent-based Modeling and Simulation on the RePast simulation platform, and the types of Agents were abstracted according to actual situation. The behavior decisions of Agents were extracted by analyzing the behavior pattern of miners in mine disaster. The escape routes of miners were simulated according to the behavior decisions of Agent in the model. The simulation results show that the model can dynamically display escaping of the miners, the success rate of Agent escaping can be improved using the escape route of the model when accident happens.

Abstract:In order to effectively improve the effectiveness and rationality of the power supplier evaluation, this study designs a supplier evaluation model based on improved fuzzy comprehensive evaluation method. First of all, based on the present situation of supplier management, we construct the index system including organization factor, product factor, supply factor, after-sale factor, and social responsibility. Secondly, in order to avoid the subjectivity of the traditional fuzzy comprehensive evaluation method, AHP+FA combination weighting method is adopted to improve the weight calculation method. Finally, the applicability of the evaluation model is verified by the example analysis, which can provide reference for the bidding management and operation decision-making of power companies.

Abstract:The two major detection methods for table tennis ball by table tennis robot's visual system are based on motion analysis or detection by single color models. The motion analysis methods have poor robustness in moving background, while the similar color and light intensity change will interfere the performance of detection methods based on single color model. A quick detection method based on multi-color models was proposed, the RGB color model and HSV color model are combined to extract the region of table tennis ball as a binary image, and the center of ball is positioned by calculating the barycenter. On this basis, Region Of Interest (ROI) technology based on previous frame position was proposed for real-time tracking. Experiments show that this method can accurate positioning table tennis ball in complex environment, the algorithm processing time is less than 10 ms, and positioning error is less than 20 mm. Its precision meets the demand of table tennis robot.

Abstract:With regard to the poor acceleration performance on GPU using the unstructured grids implicit method, this study realizes the GPU acceleration of LU-SGS implicit method based on unstructured grids with the cell-vertex scheme. With introduce the architecture of a GPU and its parallelization method, two grid reordering methods are set forth based on RCM and METIS, to improve data locality of unstructured grids and to improve acceleration performance on GPU using the unstructured grids implicit method. The ONERA M6 Wing test case is carried out to verify and validate this implementation. With two grid reordering methods, the GPU implementations achieve 63% and 69% improvements respectively. The GPU implementation obtains a speedup of 27 times compared to the CPU version running on a single core. It indicates that the proposed GPU implementation has a solid performance.