Search

Deep convolutional neural networks (CNN) have demonstrated a very powerful approach for extracting discriminative local descriptors for image description. Many related works suggest that an effective aggregation representation for deep convolutional features is particularly important in forming robust and compact image representations. In this paper, a new robust global descriptor for image retrieval is proposed by creating an effective method for aggregating local deep convolutional features weighted by regional significance and channel sensitivity through sum-pooling on multiple regions. The proposed aggregation method effectively takes advantage of multiple scales, and considers both the varied significance of regional visual content and the sparsity and intensity of response values in the channel. This can improve the ability of deep features to be described and discerned. The experimental results on six benchmark datasets demonstrate that our method can achieve retrieval results comparable to those of some popular approaches for deep feature aggregation but without fine-tuning strategies and multiple image inputs.

Question answering is an indispensable link in high school teaching. Through question answering, on the one hand, it can solve students’ learning doubts, on the other hand, it can provide teachers with teaching feedback. However, through the investigation and research, it is found that with the expansion of student size, the effect of question answering in high school is not satisfactory. This paper analyzes the current situation of question answering in high school, and designs an intelligent question answering system for high school teaching based on constructivism learning theory and cognitive structure learning theory. The system, which is the first innovative application in the field of high school teaching, integrates knowledge graph technology and intelligent question answering technology, introduces big data technology. It can solve students’ questions in time and accurately, link the knowledge points related to the questions to help students construct knowledge network graph, and the big data technology is used to analyze the students’ questioning behavior and to predict students’ learning behavior in order to feedback the teaching effect.

State-of-the-art methods on sketch classification and retrieval are based on deep convolutional neural network to learn representations. Although deep neural networks have the ability to model images with hierarchical representations by convolution kernels, they can not automatically extract the structural representations of object categories in a human-perceptible way. Furthermore, sketch images usually have large scale visual variations caused by the styles of drawing or viewpoints, which make it difficult to develop generalized representations using the fixed computational mode of convolutional kernel. In this paper, our aim is to address the problem of fixed computational mode in feature extraction process without extra supervision. We propose a novel architecture to dynamically discover the object landmarks and learn the discriminative structural representations. Our model is composed of two components: a representative landmark discovering module that localizes the key points on the object, and a category-aware representation learning module that develops the category-specific features. Specifically, we develop a structure-aware offset layer to dynamically localize the representative landmarks, which is optimized based on the category labels without extra supervision. After that, a diversity branch is introduced to extract the global discriminative features for each category. Finally, we employ a multi-task loss function to develop an end-to-end trainable architecture. At testing time, we fuse all the predictions with different number of landmarks to achieve the final results. Through extensive experiments, we compare our model with several state-of-the-art methods on two challenging datasets TU-Berlin and Sketchy for sketch classification and retrieval, and the experimental results demonstrate the effectiveness of our proposed model.

Bayesian Network (BN) can quantify the uncertain relationships among the multiple Virtual Machine (VM) related features and then predict the VM performance accurately. However, when the settings of hardware/software or the loads running on the VMs change over time, the VM-related features might be different, which will lead to the modification of VM performance prediction results. Thus, we resort to the transfer learning method and then propose a novel BN updating model, called BN_Transfer. BN_Transfer revises the weights of the updated instances constantly, and then combine the Maximum Likelihood Estimation and the hill-climbing methods to modify the parameters and structures of BN accordingly. The experiments conducted on the Alibaba published datasets and the benchmark running results on our simulated platform have shown that the BN_ Transfer can update the BN effectively as well as predict the performance of VM accurately.

At present, with the continuous development of deep learning technology, the network applied in the field of image segmentation has continuously improved in accuracy. But then the problem is that there are more and more network parameters. It is difficult to meet the needs for scenarios with poor hardware or high real-time requirements. Therefore, based on the classic DeepLab V3+ model, this paper proposes a new lightweight network structure with the goal of optimizing the size and running speed of the convolutional neural network model while ensuring the segmentation accuracy. This paper uses the MobileNet V2 structure, which is currently popular in the field of mobile terminals, to replace the original ResNet backbone network of DeepLab V3+. Then apply asymmetric convolution to improve the ASPP structure in the DeepLab V3+ network to further reduce the amount of network parameters. Finally, the Dice function and the cross-entropy function are combined to redefine the loss function of the entire network. The experimental results show that, compared with the DeepLab V3+ network, the improved network model is only 10.5% of the latter. It takes 45.8ms to process a 1220×2440 image, which is lower than traditional semantic segmentation algorithms. After testing on our self-built national costume data set, the mPA reached 86.12 and the mIoU index reached 72.83. It not only guarantees the segmentation precision, but also reduces the running time, which can better adapt to the fast segmentation network with higher requirements on segmentation performance.

Since the outbreak of COVID-19, people’s lives have been seriously affected, but with it comes another virus ’rumor’, which has the characteristics of fast-spreading, wide-spreading, and difficult to control the spreading process. Therefore, it is very important to choose an appropriate method to effectively detect COVID-19 rumors. In response to this problem, this paper crawls a data set of rumors related to COVID-19 from Snopes and construct three classes containing Fake, Real, and Unverified. We try the traditional Word Embedding Model (Word2vec, Glove, FastText) and the current Pre-training Model (BERT) to detect rumor data. In addition, this paper proposes an improved method based on the BERT pre-training model. This method obtains richer semantic information by extracting BERT hidden state output. The experiment shows that the F1 score of the method in this paper is improved, and competitive results are achieved. Our data set and source code is available at: https://github.com/byew/rumor_detection.

In order to enhance the image understanding of different regions for national costume grayscale image automatic colorization, let coloring tasks take advantage of semantic conditions, also let it can apply the human parsing semantic segmentation method to the national costume grayscale image for semantic segmentation task. This paper proposes a semantic segmentation model for context embedding based on edge perceiving. Aiming at the features of national costume grayscale image, more optimizing the model and loss function. The national costume grayscale image semantic segmentation is different from semantic segmentation of the color image, this task is more difficult for the grayscale image has no color feature. In this paper, edge information and edge consistency constraints are used to improve the national costume grayscale image coloring effect. The experimental results show that the model designed in this paper can obtain more accurate fine-grained semantic segmentation results for the national costume grayscale image.

Ethnic minority costume culture is an indispensable part of ethnic minority culture and an important content of ethnic minority culture protection and inheritance. It plays a very important role in Chinese traditional culture. The coloring of minority costume sketches has many practical application environments. It is a research topic with scientific significance and application prospects. On the basis of coloring the sketches of ethnic minority costumes on the GAN network, this paper proposes a coloring model of ethnic clothing sketches based on the Pix2Pix network, which can automatically colorize ethnic clothing sketches. The network is implemented based on the CGAN network. Among them, the ResNet is used as the network Generator. In order to achieve the constraints on the target image generation process and further ensure the coloring effect of the generated image, we use the ethnic minority costume sketch as a “label” input in the Generator, and the L1 loss is used as the loss function. The network is trained on the data set constructed in this paper. In order to verify the effectiveness of the network, we compared it with a variety of coloring methods. The results show that the peak signal-to-noise ratio reaches 24.061 and the structural similarity reaches 0.820, which further verifies that the coloring method proposed in this paper has good coloring performance.

In order to construct the knowledge graph of ethnic cultural information resources, we first use the Chinese word segmentation system and user-defined thesaurus to segment and part-of-speech tagging the data of ethnic cultural dictionary, and remove the punctuation. Then the text data is detected. If the number of continuous word segmentation is not less than the set threshold, then to perform manual word segmentation operation, and to add the result of manual word segmentation to the user-defined thesaurus of Chinese word segmentation system until there is no new word. Then we extract the attributes of the text data after the correct segmentation to build the domain knowledge graph. We detect the repeatability of the domain knowledge graph, delete the duplicate data, and store the knowledge graph. Finally, we link the stored domain knowledge graph with resources.

With the development of the Internet of Things (IoT), a new and important research direction is possible using IoT to solve the problems of information and intelligence in the metallurgical industry. This paper proposes an intelligent monitoring network based on networking technology and uses the coking process as the research object. The construction of a coking process intelligence monitoring network should focus on the formation of a perception layer network and build on a ZigBee mesh clustered network. Moreover, it also puts forward a network routing establishment and data transmission mechanism. This study provides an effective reference for the wide application of the IoT in the intelligent management and monitoring of the metallurgical process.

Feature extraction and its matching are two critical tasks in image retrieval. This paper presents a new methodology for content-based image retrieval by integrating three features, and then optimizing feature metric by diffusion process. To boost the discriminative power, the color histogram, local directional pattern, and dense SIFT features based on bag of features (BoF) are selected. Then diffusion process is applied to seek a global optimization for image matching based on fused multi-features. The diffusion process can capture the intrinsic manifold structure on a dataset, and thus enhance the overall retrieval performance significantly. Finally, a new search strategy is explored to make the diffusion process work even better when the number of retrieval images is small. In order to validate our proposed approach, four benchmark databases are used, and the results of experiments show that the proposed approach outperforms all other existing approaches.

Facial image based characterization and analysis of ethnicity, which is an important index of human demography, have become increasingly popular in the research areas of pattern recognition, computer vision, and machine learning. Many applications, such as face recognition and facial expression recognition, are affected by ethnicity information of individuals. In this study, we first create a human face database, which focuses on human ethnicity information and includes individuals from eight ethnic groups in China. This dataset can be used to conduct psychological experiments or evaluate the performance of computational algorithms. To evaluate the usefulness of this created dataset, some critical landmarks of these face images are detected and three types of features are extracted as ethnicity representations. Next, the ethnicity manifolds are learnt to demonstrate the discriminative power of the extracted features. Finally, ethnicity classifications with different popular classifiers are conducted on the constructed database, and the results indicate the effectiveness of the proposed features.

Facial semantic feature analysis for ethnical Chinese groups is one of the most significant research topics in face recognition and anthropology. In this paper, we build an ethnical Chinese face database including three ethnical groups, and then manifold learning technique is applied to analyze facial ethnic features for discriminant semantic representation. Firstly, we conduct manifold analysis on the basis of facial geometric indicators that are proposed by anthropologist, which are eventually shown not distinguishable in semantics concepts. Therefore, it is necessary to expand the scope of facial features by calculating the complete distances, angles and indexes associated with landmarks. Then, mRMR-based feature selection is applied to select 2926 distance indicators, more than 210,000 angle indicators and more than 4,100,000 index indicators ethnical feature representation, and 5 datasets with features of distance, angle, index, anthropology and combinations are obtained. Secondly, several popular manifold learning methods, such as LPP, ISOMAP, LE, PCA and LDA are utilized to investigate the ethnic features obtained above, and the results show the distinguishable manifold structure of facial ethnical features and clusters in 4 of the 5 datasets. In order to evaluate the validity of filtered features, the classification algorithms, J48, SVM, RBF Network, Bayesian, and Bayes Network in Weka, are carried out based on the filtered features. The experimental results reveal that the average of classification accuracy on the dataset with combined features is higher than other datasets, and the corresponding indexes are more salient than other geometric features. Finally, the sub-manifold structures with semantic concepts are found based on the ethnic facial data. Facial features of three Chinese ethnic groups exist in different ethnic semantic sub-manifolds in the low-dimensional space. Facial measurement indicators obtained by manifold analysis and feature selection provide not only a method for computational facial ethnic groups analysis, but also an enrichment and improvement to the related research in anthropology.

In normal computer networks, the data transmission between two sites go through the shortest path between two corresponding vertices. However, in the setting of software definition network (SDN), it should monitor the network traffic flow in each site and channel timely, and the data transmission path between two sites in SDN should consider the congestion in current networks. Hence, the difference of available data transmission theory between normal computer network and software definition network is that we should consider the prohibit graph structures in SDN, and these forbidden subgraphs represent the sites and channels in which data can’t be passed by the serious congestion. Inspired by theoretical analysis of an available data transmission in SDN, we consider some computational problems from the perspective of the graph theory. Several results determined in the paper imply the sufficient conditions of data transmission in SDN in the various graph settings.

At present, the consistency detection of products mainly relies on manual method which has defects such as difficulty in identifying similar categories and low efficiency. Recently, the methods which are based on image classification have been widely used in it. However, image classification is based merely on simple visual features without making full use of other categrizable attributes. The approach this paper adopt to solve the problem is a joint detection algorithm which is based on deep learning. This algorithm locates product areas and implements image segmentation according to the default information code. Then it takes the segmented image as input of convolutional neural network model for classification. A reflection-assisted decision-making algorithm is used to solve the defect that the approximate class output probability is unstable. Finally, the consistency judgment is made according to the extracted default class information and product category. The experimental results obtained in this research show that the proposed algorithm can segment the products image accurately and quickly. The joint decision-making algorithm has better accuracy and stability than the single classification network.

This paper uses deep learning related algorithms to classify and recognize the ethnic costume images of the Wa and Yi people. In the deep learning framework Tensorflow, the deep convolution network VGG model was migrated to the ethnic costume recognition task, and the image recognition of ethnic costume based on VGG was realized. By adjusting the size of the convolution kernel and the number of convolution layers of the VGGNet network model, a neural network model of ethnic costume image recognition suitable for the classification and recognition of the Wa and Yi people is constructed. By training the images in the ethnic costume image library, iteratively adjusting the parameters of Batch_size, Epoch, Dropout and other parameters of the network model, the comparative experiments are on, and the classification recognition rate of ethnic costume images under different network model parameters is analyzed.

Under the background of “Internet + education”, the information technology is changing rapidly, and the mobile learning technology is becoming more and more popular. In recent years, learning based on mobile terminal device is a popular way, and people like to use their fragmented time for learning. Because people depend on the mobile Internet at present, the way of life and learning has also changed. The maturity of artificial intelligence technology has brought great convenience to people's life. It is the requirement of the times to apply artificial intelligence to education field and promote the development of education. As a product of the development of mobile Internet, WeChat has become an indispensable part of people's communication tools. Almost everyone in contemporary youth uses WeChat. As a light application based on the secondary development of the WeChat platform, WeChat Mini Program is also a boom in the development of the times. Combining the advantages of the development of WeChat Mini Program and the development trend of the mobile learning technology, this paper mainly analyzes the current situation of secondary school students participation in extra-curricular tutoring, expounds the advantages of WeChat Mini Program as a mobile learning platform, combines artificial intelligence technology with WeChat Mini Program platform, and constructs a homework tutoring model for secondary school students based on WeChat Mini Program as a classroom teaching assistant for secondary school students. It is hoped that the study of WeChat Mini Program as a homework guidance platform can provide a reference to improve students' ability to think and solve problems.

In most real-world document collections, there are various types of labels that usually carry context information, such as label hierarchies or textual descriptions. Nonetheless, the commonly-used approaches to modeling text corpora ignore this information. Label embedding can reflect more extensive label context information and have a capability of leveraging various sources of information. In this paper, we propose a multi-label topic model conditioned on label embedding, which incorporate label embedding into the generative process of multi-label topic models in text domain, so as to improve documents classification accuracy and topic quality. By introducing Dirichlet-multinomial regression (DMR) framework into a multi-label topic model called Labeled LDA(LLDA), our model apply an exponential priori constructed previously with label embedding on the hyperparameters of document-label distribution, which reflects the effects of label embedding on label probability distribution. The experimental results demonstrate the potential of our model through an exploration of a standard document dataset.

The superpixel generated by clustering of simple linear iterative clustering (SLIC) algorithm is prone to under-segmentation problem in image detail, and the image boundary segmentation is not accurate. We propose an algorithm to perform motion clustering on the boundary pixels that initialize the uniform segmentation region until the convergence produces superpixels, which improves the accuracy of clustering in the iterative process and reduces the phenomenon of mis-segmentation. The experimental results show that the image boundary segmentation algorithm is better than the SLIC algorithm.

KNN (K-nearest neighbor) is a simple and practical classification algorithm, but it is less efficient when dealing with massive high-dimensional data. Parallel computing is an effective way to accelerate big data calculations. In order to improve the efficiency of KNN, an optimization method based on GPU parallel computing is proposed in this paper. In the distance calculation stage, the parallelism is increased to the numerical value, and the independent components are calculated in parallel before the thread collaborative calculation. In the distance sorting stage, a method of judging order is proposed. This method based on shared memory gives Odd-even sorting the parallel ability to determine whether the sequence is ordered. The experimental results obtained in this research show that the proposed method can obviously improve the execution efficiency of KNN algorithm in high-dimensional samples and partial ordered distribution.

Minority ultra-high definition video resources with ultra-high resolution will cause the occupation of large storage space, so the mainstream network rate is difficult to satisfy the requirement for smooth playback. To address the above problem, we first adopt distributed storage technology to resampling and splitting the ultra-high definition video. Then, multi-level dynamic image frame and resolution archiving are established and combined with image blocks information to form dynamic image playback archiving. This method effectively processes the conflict between ultra-high-definition video and the upload bandwidth of the server network, and thus solves a kind of ultra-high definition video broadcast problem with little consistency between frames and the need for local amplification of images. Moreover, it provides a way of broadcasting and displaying for the inheritance and protection of minority culture.

The Internet of Things (IoT) is regarded as the third wave of information industry in the world after computer Internet and mobile communication network. It has attracted worldwide attention recently due to its great application prospect, and become an important research hotspot in the field of information technology at home and abroad. Localization is a very critical issue of IoT. It is an important supporting technology for most applications of IoT. In this paper, a novel intelligent computing based localization algorithm for IoT is presented. It combines the advantages of invasive weed optimization (IWO) and simplified quadratic approximation (SQA) to achieve better positioning performance. The simulation results show that, the presented localization algorithm has higher localization accuracy compared with other similar algorithms. Moreover, it has a strong robustness for communication ranging error, and shows higher practicability.

In this paper, we propose a novel color image retrieval approach by using an effective fusion of two types of histograms extracted from color and local directional pattern (LDP), respectively. First, we describe the extraction process of color histogram and LDP. Secondly we present these two features and then develop an effective fusion procedure including feature normalization and a new similarity metric. Thirdly, this new approach is validated after extensive comparisons with several existing state of the art approaches on two benchmark datasets including the Wang’s dataset and large size of the Corel-10000 dataset. Finally, a friendly interface for this proposed retrieval system is designed and used to show some retrieval results.

In this paper, we present an automatic coloring model for ethnic costume sketches based on a generative adversarial networks. The proposed model is composed of a 6-layer U-net structure generator and a discriminator with 5-layer convolutional neural network. Then the loss function and the weights of true reliability value in the discriminator are optimized. And, finally the constructed sketch database is used for training to get an automatic coloring model. The results of a large number of sketch coloring experiments show that the model has good learning ability on exploring color law of ethnic costumes and can achieve better coloring effects.

Urban road image segmentation is an important technology of intelligent city management and pilotless driving. In order to improve the effect of image segmentation, direct at the deficiency of single seed point and fixed threshold of traditional region growing algorithm, a seed selection method based on the gray level of two-dimensional histogram and local variance is proposed, and the dynamic threshold is used to change the region growing rule. The experimental results show that the seeds selected by this method can be highly representative, and realize the complete segmentation of the image. Based on the dynamic threshold region growth rule, the image segmentation has a better effect.

In recent years, along with our country education constantly reforms and develops, the education for nationalities research is also progress and obtained good results constantly. This article utilized the method of knowledge map, keywords co-occurrence analysis to our country in recent years, analyses the education for nationalities situation, constructed our country education for nationalities research area keyword knowledge map and development facility knowledge map and researcher knowledge map, show and revealed the focus and front of our country education for nationalities research in recent years, reflected the recent years Our country education for nationalities area research leader.

Multi-label classification with many labels are common in real-world application. However, traditional multi-label classifiers often become computationally inefficient for hundreds or even thousands of labels. Therefore, the label space dimension reduction is designed to address this problem. In this paper, the existing studies of label space dimension reduction are summarized; especially, these studies were classified into two categories: label space dimension reduction based on transformed labels and label subset; meanwhile, we analyze the studies belonging to each type and give the experimental comparison of two typical LSDR algorithms. To the best of our knowledge, this is the first effort to review the development of label space dimension reduction.

New word discovery possesses a significance in NLP. This paper first reduces noise to the corpus of micro-blog and employ the new filtering algorithm to filter the candidate words, then improves the traditional mutual information and adjacency entropy method respectively and put forward enhancement of mutual information and relative adjacency entropy. In terms of multi-feature massive data generated by a large-scale corpus to recognize the new words, the MapReduce parallel computing model is exploited to extract three features such as, enhancement of mutual information, relative adjacency entropy and background document frequency, to improve the parallelization. With the extracted three features, the feature vectors of the candidate words are formed, and a SVM model can be trained by training the labelled corpus. The experiments show that the proposed method shortens the time required by the whole recognition process. In addition, compared with the existing methods, the F-value reaches 86.98%.

As a NP-hard combinatorial problem, nurse scheduling problem (NSP) is a well-known personnel scheduling task whose goal is to create a nurse schedule under a series of hard and soft constraints in a practical world. In this paper, a variant of structure-redesigned-based bacterial foraging optimization (SRBFO) with a dynamic topology structure (SRBFO-DN) is employed for solving nurse scheduling problem (NSP). In SRBFO-DN, each bacterium achieves cooperation by information exchange mechanism switching the topology structure between star topology and ring topology. A special encoding operation of bacteria in SRBFO-DN is adopted to transform position vectors into feasible solutions, which can make SRBFO-DN successfully dealing with this typical difficult and discrete NSP. Experiment results obtained by SRBFO-DN compared with SRBFO and SPSO demonstrated that the efficiency of the proposed SRBFO-DN algorithm is better than other two algorithms for dealing with NSP.

Describing and representing multi-source and heterogeneous knowledge is a popular research topic in recent years. After investigating knowledge forming process based on multi-source heterogeneous information resources, we present a new approach in which different information resources are put into a mutual RDF(S) data model, and semantic reasoning of RDF(S) is conducted. Moreover, a knowledge base construction framework for multi-source heterogeneous information source with combination of Ontology knowledge model is put forward, and an algorithm of knowledge base construction is also proposed, in which the core issues are knowledge inclusion and updating. Then the time complexity of our algorithm is analyzed. Finally, in order to solve the heterogeneous, and uneven horizontal of geographical distribution of ethnic minority information resources in Yunnan Province, we use the proposed method to construct a domain knowledge base for ethnic minority information resources, and use this model to evaluate the efficiency for the knowledge inclusion algorithm in responding time and indexing responding time for different data resources in our experiments.

In order to improve energy efficiency and decrease network delay in wireless sensor network applied to emergent event monitoring, a new data gathering algorithm based on mobile agent and event-driven is proposed for cluster-based wireless sensor network. The process of dynamically clustering the sensor nodes is based on the event severity degree, by which the scale and lifetime of clusters are determined. And a multi-hop virtual cluster is formed between the base station and the cluster heads in which the base station is regarded as its cluster head. The order of nodes visited along the route by mobile agent has a significant impact on the algorithm efficiency and the lifetime for wireless sensor network. In this paper, the next hop in route planning for mobile agents is determined by the residual energy, the path loss and the stimulated intensity. The mobile agents can gather information by traversing all member nodes. The theory analysis and simulation results show that mobile-agent-based model has a better performance in energy consumption and network delay compared to C/S model. And mobile agent is more suitable for wireless sensor network than C/S model in data aggregation. Furthermore, DGMA will provide a more network applied to a large scale emergent event monitoring.

Wireless Sensor Network (WSN) & Radio Frequency Identification (RFID) integrated technology (Wireless Sensing Identification, WSID) is an inevitable trend in WSN and RFID fields in the future, Especially the RFID based on automated identification helps the objects identification in traffic management to become more simple and exact. In this paper, RFID is integrated into highway traffic monitoring system architecture based on wireless sensor network, thus wireless sensing identification traffic monitoring network (WSTIMN) is presented. And its network architecture, hardware and software structure of smart nodes, network protocol stacks are proposed and described in detail. Based on the above structures, we implemented the prototype system in the experimental platform. The experimental results demonstrated that the integration and amalgamation of RFID and WSN not only effectively expanded and improved the original system function, simplified some complicated function operations, but also improved those two major technology itself.