Indoor localization is a prerequisite for Location Based Services(LBS),it is research hotspot in recent years,and receive widespread attention from academia and industry.With the popularity of smart terminal,many kinds of indoor localization method which is based on WiFi and smart terminal are proposed.Compared with the previous localization method,indoor localization based on WiFi and smart terminal has advantages.Through the key factors of large scale application of indoor localization,this paper analyzes and summarizes the basic methods of indoor localization based on WiFi and smart terminal systematic.With the comparison of typical work in recent years,it points out the shortage of present work,and prospects on the difficulties and direction of the research on indoor localization based on WiFi and smart terminal.

Multi-path TCP(MPTCP) and Concurrent Multipath Transfer for SCTP(CMT-SCTP) are two technologies of the concurrent multi-path transfer for NGI,which are hot topics. Compared with the traditional TCP and SCTP,the congestion control of multi-path transport protocols will face the problem of unfairness. This paper begins with an overview of the traditional TCP congestion control algorithm,and then in view of the unfairness of concurrent multi-path transfer,basing on the resource pooling method,it introduces the slow start threshold ratio factor,increase factor,per-flow aggressiveness factor,deduces three new algorithms to solve the problem of unfairness,and carryies out the contrastive analysis through simulation.

Many dynamic binary translators chose basic block as the unit of translation and execution, the basic blocks to be translated in dynamic binary translation may have the same instructions as the other ones which have been translated. This situation consequently leads to repeating translation and incurs the unnecessary overhead. This paper focus on how to detect and eliminate the overhead caused by basic blocks overlapped redundancies from the view of dynamic binary translation optimization technology. Experiment results show there are about 5% overlapped redundancies in the course of translation. By eliminating the redundancies results in a speed up of 1%~4% in translation performance.

A radix-4 Montgomery modular multiplication and the optimized circuit architecture are presented. It can reduce iterations of traditional radix-2 modular multiplication to about 50%. Based on this module, the implementation of high speed RSA encryption processor follows completely parallel modular exponentiation flow with Carry Save Addition(CSA) structure to perform long integer arithmetic. This avoids the repeated interim output/output format conversion. Result shows that the optimization is technology independent and thus should suit well for not only FPGA implementation but also ASIC. This design can complete a standard 1 024 bit RSA encrypt operation with only 9 836 clock cycles. Compared to the recently proposed design in the literature, the proposed design can achieve an increase of over 50% in throughput.

This paper improves traditional collaborative filtering algorithm, proposes a new user profile based on concept hierarchy tree, which can make recommender systems still work even when users have no common rating items. In the process of similarity calculation and recommendation formation, it also uses concept hierarchy thought to generate recommendation lists by different categories, avoiding recommendation lack of diversity. Experimental results on MovieLens dataset show that the improved algorithm can provide better prediction in either accuracy or diversity aspect.

Latent Dirichlet Allocation(LDA) is presented to express the distributed probability of words. The topic keywords are extracted according to Shannon information. Words which are not distinctly in the analyzed text can be included to express the topics with the help of word clustering of background and topic words association. The topic meaning is attempted to dig out. Fast Gibbs is used to estimate the parameters. Experiments show that Fast Gibbs is 5 times faster than Gibbs and the precision is satisfactory, which shows the approach is efficient.

In order to overcome the limitation of literal matching and lacking semantic concept of the traditional keyword extraction algorithm, this paper presents a Semantic-based Keyword Extraction(SKE) algorithm for Chinese text. It uses semantic feature in the keyword extraction process and constructs word semantic similarity network and uses betweenness centrality density. Experimental results show that compared with the statistic based keyword extraction algorithm, the keywords SKE algorithm extracted are more reasonable and can represent more information of the document’s topic, and the SKE algorithm has a better performance.

In order to research the influence of information characteristis on information spreading in social network,this paper proposes an information spreading model which combines information characteristics with memory effects.The information characteristics are measured by the attributes of information amount and information adhesion.It performs the model on regular network,small-world network,random network and BA scale-free network.Results indicate that in the speed and scope of information spreading,the attribute of information amount is more effective than the attribute of information adhesion.Under certain information characteristic,it exists a fixed value when individuals select spread information.Especially,on the regular network,the probability of spreading information can reach the maximum when individuals receive information at the second time.

Verilog’s formal semantics using the abstract states machine are studied, and the formal definition of assignment statements and delay/event control mechanism is given. Comparing with Borger’s definition on VHDL, the key methods on how to translate Verilog description to VHDL are explained. In the end, the simulation differences before and after translation are studied.

In order to figure out how memory faults affect high-performance servers, this paper presents a multi-layer fault injection method for Itanium architecture computers. HMFI(Hierarchical Memory Fault Injector) is designed and developed to inject memory faults in physical level, operating system kernel level and process level. Experimental results demonstrate the effectiveness of HMFI in evaluating the fault-tolerant properties of complex computer system.

Aiming at the factors of resource constraints that exist in the process of course schedule, this paper proposes an algorithm combining Genetic Algorithm(GA) and constraint satisfaction algorithm to solve course schedule problem. Course schedule tasks are sorted with constraint satisfaction algorithm, and a single course schedule task’s timetable is allocated and optimized with GA. In this algorithm, the result of single course schedule task is global optimal. Experimental results show that this method is feasible to improve the performance and the efficiency.

A new architecture of parallel BCH encoder and decoder applied in NAND Flash Controller is proposed. In order to obviously increase the throughput of decoder, pipeline operation and prefetch decoding in group operation are applied in the design. It takes 565 cycles to correct 8 bit random error after NAND Flash’s 2 KB page read operation, which is a quarter of the time cost by prefetch & decode in page.

By combining several parameters selection approaches of Support Vector Machine(SVM), this paper proposes a method that defines parameters directly by analyzing training samples. Experimental results based on several standard test data sets show that the method achieves better prediction accuracy and consumes less time compared to traditional grid search methods. It provides one way to deal with the problem of selecting parameters and time consuming in application of SVM.

Aiming at the challenges of single-text representation and low accuracy of existing video content text-representation models, a video content text-reprsentation model that integrates frame-level image and audio information is proposed.The network structure of the model includes a single-mode embedding layer based on a self attention mechanism, and learns single-mode feature parameters.Two schemes, joint-representation and cooperative-representation, are adopted to fuse high-dimensional feature vectors output from the single-mode embedding layer, so that the model can focus on different objects in the video and their interaction, thereby generating richer and more accurate video text representation.The model is pretrained through large-scale datasets, and representation information, such as video frames and audio carried by the video, are extracted and sent to the coder to realize the text representation of the video content.The experimental results on MSR-VTT and LSMDC datasets show that the BLEU4, METEOR, ROUGEL, and CIDEr scores of the proposed model are 0.386, 0.250, 0.609 and 0.463 respectively.Compared with the model released by the IIT DeIhi in the MSR-VTT challenge, the proposed model improves the indexes above by 0.082, 0.037, 0.115 and 0.257 respectively.The model in this study can effectively improve the accuracy of the video content text-representation model.

This paper interprets the concepts of coreference resolution, and introduces the state-of-the-arts in coreference resolution from three aspects: corpus resources, evaluation measures and resolution algorithms. It analyzes the critical problems of coreference resolution: automatic capture and presentation of structured syntactic knowledge, automatic capture and usage of underlying semantic, cross-document coreference resolution. Based on the analysis, it explores its international research trend.

DNS is a critical component of the operation of Internet applications. The Internet is greatly affected if DNS is attacked. DNS spoofing is one of the most popular attack means with the character of high dormancy and good attack effection. But so far, little is done to defend the systerm against this attack. Three methods are presented to detect DNS spoofing attack, and then another three techniques are proposed to identify the bogus packets and the right ones to ensure DNS service even attacked.

Most existing algorithms for document layout analysis are sensitive to the parameters and have weak applicability. In order to make up these deficiencies, this paper presents an algorithm of region formation based on Support Vector Machine(SVM) for analyzing Chinese document. In the proposed technique, it selects connected components called seed connected components which can be used to represent regions as the first feature for training. SVM is employed to form region since it has strong learning and generalization ability. The technique decides the reading order by exploiting the projection method. Extensive experimental results show that the proposed algorithm is more effective to analyze different kinds of document layout than the state-of-the-art methods.

A genetic algorithm to handle constrained optimization problem is proposed. This method searches the decision space of a problem through the arithmetic crossover of feasible and infeasible solutions, and performs a selection on feasible and infeasible populations respectively according to fitness and constraint violation. It uses the boundary mutation on feasible solutions and the non-uniform mutation on infeasible solutions because the solutions usually deviate from the constraint domain after the traditional mutation operation. It maintains the population diversity through dimension mutation. Numerical results show that it is an effective algorithm.

To raise the measurement accuracy of binocular stereo vision system, working principle of binocular vision system and error model in all aspects are analyzed. The calibration error and the image rectification error can be reduced by improving the accuracy of calibration template and increasing calibration images. By analyzing image recognition errors’ influence on the measurement accuracy of the visual system, the relationship between structural parameters of visual system and measurement errors is introduced. The influence of structural parameters to visual system is determined. The influence of object distance on measurement accuracy is discussed and a proper object distance scope is given.

3D face reconstruction algorithms require more than one image to realize reconstruction and their efficiency is low. Aiming at the problems, this paper proposes a 3D face reconstruction algorithm based on single frontal photo. It uses Thin Plate Spline(TPS) function to build 3D face model by establishing the point-to-point correspondence among 3D faces in the database, so that an average 3D face model is built. It uses Levenberg-Marquardt Algorithm(LMA) to optimize the shape coefficients. The input image is projected orthogonally to the 3D shape to generate the face texture. Experimental results show that the method can generate realistic 3D faces efficiently.

In order to increase the positioning accuracy of indoor positioning system based on Received Signal Strength Indicator(RSSI), a Bayesian indoor location algorithm based on RSSI is proposed. The algorithm preprocesses the RSSI signal by Gaussian filtering, uses the triangle centroid algorithm to calculate the unknown node’s initial coordinates, and handles it through Bayesian filtering to get the coordinates that is more accurate. Experimental results show that the precision of proposed algorithm can reach 98%, which reduces the positioning error.

This paper analyzes Chinese character structure feature and evolution, combines the character recognition process together on the basis of Chinese character database and changing in recognition target, researches and comprises the advantages, disadvantages and the application range of five different kinds of Classification recognition method such as Bayesian method, quadratic discriminate function approach, Markov model, neural network classifier and Support Vector Machine(SVM). Results shows that Chinese character recognition method including recognition should be improved on the aspects of target, feature extracting and train set preparing.

A new reciprocal algorithm is provided to realize the traditional reciprocal division on Field Programmable Gate Array(FPGA) with high precision and fast speed. In order to solve the reciprocal of the divisor, the Taylor series expansion is used combining with searching approximation to get a fitting binomial of degree 1 with one variable in every separated section, and the Newton-Raphson iteration is used one time to improve precision. The result of timing simulation proved that the divider based on this new division algorithm can be easily realized on FPGA. The divider can get the answer of 2-34 available accuracy only six clock pulses after input variables being given, and the highest operating frequency is 86.95 MHz.

The traditional k-means algorithm has sensitivity to the initial start center. To solve this problem, a new method is proposed to find the initial start center. First it computes the density of the area where the data object belongs to; then finds k data objects all of which are belong to high density area and the most far away to each other, using these k data objects as the initial start centers. Experiments on the standard database UCI show that the proposed method can produce a high purity clustering result and eliminate the sensitivity to the initial start centers.

Primitive elements and primitive polynomial play very important roles in the theory of the finite field. It is the premise of solving the problem about code sequences and searching the primitive elements can come down to searching primitive polynomial. This paper gives a new algorithm for searching primitive polynomials in the binary field making use of the algorithm for searching the minimal polynomial, and also gives out the minimal polynomial in the searching process. It shows the efficiency analysis of the algorithm.

Because of the accumulation of positioning errors due to the drift of Personal Dead Reckoning(PDR) for indoor localization, this paper puts forward a Micro Electro Mechanical System(MEMS) sensor integration of PDR algorithm. It includes the following three points: It obtains displacement information based on typical pedometer principle and stride length estimate, joins dynamic time window and acceleration threshold in stride calculation algorithm, and gets the angle between the walking direction and magnetic north direction with the magnetometer and gyroscope to correct the positioning accumulative error. It realizes indoor location and navigation. The result proves that the distance errors of indoor localization experiment can be controlled within 5%, and the stride calculation is highly precise, the azimuth algorithm is easy to deal with, and it can rectify error because of long time drift in certain extent.

Applying ElGamal signature scheme to the elliptic curve cryptosystem may introduce an undefined operation of point multiplex in elliptic curve. In order to solve the problem, this paper improves processes of signature generation and validation, using a simple algebra operation instead of multiplication operation, and presents feasibility and security analysis. Menezes Vanstone(MV) algorithm is improved to reduce its data expand rate. Experimental results show that its operation speed is faster than RSA or ECC-E. Execution efficiency and key length of the improved algorithms make them more efficient in the application of smart card computation.

A KNN classification algorithm based on k-nearest neighbor graph for small sample sets is presented to improve the classification accuracy, which partitions the k-nearest neighbor graph into clusters with high similarity, labels the unlabel data of each cluster with the label of the label data in the same cluster, and deletes the noise data. The sample set is expended by this method. The algorithm use the expended sample set to label the unlabel data. The presented algorithm is demonstrated with standard datasets, and the experimental results show the algorithm can enhance the accuracy of classification, reduce the influence of the value of k, and achieve a satisfying result.

This paper introduces the result of a test that evaluates the effectiveness of Compute Unified Device Architecture(CUDA) using NVDIA GeForce8800GT and the compiler Visual Studio 2008. It tests the speed of NVIDIA CUBLAS, CUDA kernel, common C program, Intel MKL BLAS, CUDA driver API program, FFTW and CUFFT Library in matrix-multiplication and Fast Fourier Transform(FFT). Test result of the large scale data shows that the computing ability of GPU is 25 times better than that of CPU.

Bicubic interpolation is an effective way to get well-qualified high-resolution image, but with high calculation load. The common used interpolation methods are discussed. A new region-based bicubic interpolation is proposed. Without segmenting the image, the mean value of the four neighboring points of the interpolated point is calculated which is used to divide the image into two regions: the flat region and the complex region with more details. Two different interpolating algorithms are chosen for each region. Experimental results show that the proposed algorithm can keep the image quality equal to the original algorithm while reducing the calculation load more than 10 percents. It is useful for applications.

The mass data from Web are distributed, heterogeneous and dynamic, so the current data mining system based on single node has developed to a bottleneck. Using the advantage of cloud computing——distributed processing and virtualization, this paper presents a Weblog analysis platform under the Hadoop’s cluster framework based on cloud computing, it also presents a hybrid algorithm which can distributed process in the cloud computing environment. To further verify the effectiveness and efficiency of the platform, it uses the improved algorithm to mine users’ preferred access path in Weblog on the platform. Experimental results show that, using distributed algorithm to process large number of Weblog files in the cluster, can significantly improve the efficiency of Web data mining.

Modeling and predicting of network traffic are important for internet management and security prediction. Considering about draggle disposal and bad service quality, different kinds of classic traffic prediction methods are reviewed. Comparison about their suitable data, modeling complexity, prediction precision and suitable scene are presented. Experimental results prove that prediction model should correlate to scene tightly. It need to choose different models according to the traffic characters and prediction target in practices.

As applications programs based on MapReduce model growing up, Hadoop’s performance depends on applications. From applications, this paper analyzes the limitations and shortcomings of Hadoop architecture and storage technology, and proposes a feasible and cost-effective solution to address the problem. It builds a series tests to convince the idea. The solution is multi-level parallel, both on job and I/O levels, which makes full use of disk and network bandwidth, reduces I/O bottlenecks and improves performance.

The self-organizing maps (SOM) is an unsupervised learning algorithm, which is capable of self organization and visualization and has been used in many areas. In this paper, the background of self-organizing maps algorithm and basic algorithm are introduced, the parameter setting-up and disadvantage of SOM are analyzed. The improved algorithms have been reviewed emphatically, and research hotspot and applying fields are also discussed briefly. Finally, some remarks on future research are given.

Yao’s Millionaires’ problem is a typical problem of secure multi-party computation, but most solutions are inefficient. Based on commutative encryption scheme, this paper proposes an efficient and secure solution to millionaires’ problem, which reduces the problem to the set- intersection problem by 0-encoding and 1-encoding for private inputs. Proof of security is followed. There is no complicated modular exponentiation in this solution which only needs O(n) encryption/decryption and 4 rounds of communication. It is more efficient than other solutions.

Latent Dirichlet Allocation(LDA) model can be used for identifying topic information from large-scale document set, but the effect is not ideal for short text such as microblog. This paper proposes a microblog user model based on LDA, which divides microblog based on user and represents each user with their posted microbolgs. Thus, the standard three layers in LDA model by document-topic-word becomes a user model by user-topic-word. The model is applied to user recommendation. Experiment on real data set shows that the new provided method has a better effect. With a proper topic number, the performance is improved by nearly 10%.

Two one-class classification models, one-class support vector machine and support vector data description, which are based on support vector machine and one-class classification are introduced. The internal relationship and parameters optimization of the two models are also analysed, and the exists of disadvantages and improvements of the two one-class classifiers are summarized.

Geometric feature extraction from image is a crucial step in image classification and recognition algorithms. It requires to calculate the Minimum Enclosing Rectangle(MER) of the image to get some attributes, such as length and width. This paper proposes an algorithm based on the combination of the vertex chain code and the discrete Green theorem to extract MER from image. According to this algorithm, it can calculate the area, centroid, and principal axis of the image easily, only with its point coordinates from the veritical or hrizontal direction of the vertex chain code. Experimental results show that the rotation method’s speed is two times of the current algorithms, while the principal axis method is three times of the rotation method.

Based on analyzing the shortages of K-Nearest Neighbor(KNN) algorithm in solving classification problems on imbalanced dataset, a novel KNN approach based on weight strategy(GAK-KNN) is presented. The key of GAK-KNN lies on defining a new weight assignment model, which can fully take into account the adverse effects caused by the uneven distribution of training sample between classes and within classes. The specific steps are as follows: use K-means algorithm based on Genetic Algorithm(GA) to cluster the training sample set, compute the weight for each training sample in accordance to the clustering results and weight assignment model, use the improved KNN algorithm to classify the test samples. GAK-KNN algorithm can significantly improve the identification rate of the minority samples and overall classification performance. Theoretical analysis and comprehensive experimental results on the UCI dataset con?rm the claims.

To deal with the new time and space challenges of the machine learning problem algorithms from large scale data, this paper focuses on sparse-learning and categorizes the L1 regularized problem’s the-state-of-the-art solvers from the view of multi-stage and multi-step optimization schemes. It compares the algorithms’ convergence properties, time and space cost and the sparsity of these solvers. The analysis shows that those algorithms sufficiently exploiting the machine learning problem’s specific structure obtain better sparsity as well as faster convergence rate.

Different applications have different security sensitivities, which are determined by subjective judgment in the process of security research on Internet of Things(IOT). This paper presents a Security Model of IOT Based on Hierarchy(BHSM-IOT), which can be used to analyze the topology structure, predict the attack source and attack type and determine the application level of security domain, so that appropriate safety technical configuration can be made. The model is applied in a wisdom campus system of a university, and result proves that it is helpful for the managers to construct a more safe and stability system for the students and teachers with the model.

According to the features of Message Passing Interface(MPI), the cloud computing application methods based on MPI, including the MPI cloud computing algorithm design model, cloud computing principles, the core model, and the process are proposed. The distributed characteristic and parallel characteristic are introduced. Theoretical analysis results show this algorithm is feasible, effective and superior to the traditional parallel technology, and it can provide the new method to distributed and parallelize the ordinary algorithms.

Aiming at similarity and instability of gesture activity signal in gesture interaction, a gesture recognition scheme based on three-axial accelerometer is presented. It utilizes a MMA7260 accelerometer to capture the acceleration signal of dominant wrist. An activity detection algorithm is used to auto determine the data stream which containing interesting motion according to the features of gesture activity signal. After denoising and resampling the acceleration data streams, gesture features are extracted, and Discrete Hidden Markov Model(DHMM) is built for gesture recognition. Experimental results demonstrate the effectiveness of the scheme.

In order to realize landmark tracking by Unmanned Aerial Vehicle(UAV), this paper embeds online feature selection into particle filtering algorithm, and adopts adaptive transition model. Candidate feature set is composed of linear combination of R, G and B pixel values. Histograms of feature values for pixels on the object and in the background are computed for obtaining log likelihood ratio and variance ratio. The feature with the best discrimination is selected for computing the likelihood image. Two sets of particles are obtained via different approaches for estimating the position of the object. Experimental results show that the algorithm provides more reliable results and it is more robust.

This paper conducts a survey of pricing models, relevance match algorithms, and effective statistical models for computing advertising, analyzes and compares these approaches, like Cosine, Okapi BM25, feature learning, hierarchy-learning and Multinomial language model, and conclusively points out the feasible improvement and future of research in this field.

A method for generating the depth image from multi-view images based on color segmentation is presented. Mean-shift algorithm is used to extract homogenous regions in the reference image. The initial disparity maps applying local window-based matching algorithm is got. According to certain criteria, these initial depth maps are merged into one disparity map to improve the quality of disparity map, and optimized the disparity map by post processing. The disparity map is converted to a depth map. The algorithm can deal with the occlusion problem effectively, which increases the matching accuracy, and improves the quality of the disparity map.

This paper analyzes the current question parsing techniques in Question Answering(QA), and proposes a question feature exaction method based on syntactic structure and semantic information. The method parses question trunks based on syntax structure, then extracts characteristics of question from the semantic information point of view and uses modified Bayesian algorithm in question classification. Experiments show that this method can raise the accuracy of the question classification.

This paper analyzes the up-down layer topology of multi-source media streaming network, and presents a topology optimization strategy based on loading balance. This strategy makes use of the abundant network resource by P2P technology and combines media servers function with group broadcasting theory. It can make all media streaming network’s topology balanced, and bring on all system’s loading balanced. Playing delay is reduced, and system stability is enhanced when system expansibility is ensured.

Existing watershed algorithm is sensitive to noise and has over-segmentation problem. Based on previous studies and combining with the current progress, this paper introduces watershed algorithm simulating flood and watershed algorithm simulating rain. It sums up four types of improved methods for watershed algorithm according to defects of the traditional algorithms. Problems and further research aspects of improved methods for watershed algorithm are discussed.

According to the problem that high-speed processors can not be emulated through common JTAG-interface in a long distance, enhanced JTAG interface is proposed. Several factors are analyzed quantitatively that relate to the failure of emulating the high-speed processors in a long distance, through the discussion of character and timing of JTAG-interface-signals. According to the conclusion, drivers and compensation of signal integrity are given to the signals of common JTAG interface. The circuit and test result of the design is also discussed. The results show high-speed processors can be emulated through enhanced JTAG-interface over 120 cm.