Abstract:Multiple kernel learning shows good superiority in solving irregular and large-scale data problems. Regularization path is a method to select the optimal model by solving the multiple kernel learning multiple times.Aiming at the problems that the kernel matrix size is large， the computational cost is high and the efficiency of the optimization model is affected when multiple kernel learning regularization path processes large-scale data， a multiple kernel learning regularization path approximation algorithm based on CUR matrix decomposition is proposed, which is named MKLRPCUR.This algorithm firstly adopts CUR algorithm to obtain multiple decomposition matrices of low-rank opproximation matrix of kernel matrix.Then， in the solution process， the low-dimensional decomposition matrices are used to replace the kernel matrix， and the order of the correlation matrix calculation is adjusted， thereby simplifying the calculation of the kernel matrix and the Lagrange multiplier vector product.MKLRPCUR algorithm reduces the calculation scale of matrix， optimizes matrix calculation， and improves the calculation efficiency of exact algorithm.The relative error of the low-rank approximation matrix and the time complexity of the algorithm are theoretically analyzed to verify the rationality of the approximation algorithm.At the same time， the experimental results on the UCI dataset， ORL and COIL image databases show that the proposed approximate algorithm not only ensures the accuracy of learning， but also reduces the running time of the algorithm and improves the efficiency of the model.

Abstract:Using hierarchical structure to classify image set which has the object labels identified by images is one of the important research issues in image automation classification. The previous researches have already implemented the hierarchical structure construction for the labeled images, and now there are only a few methods to consider the influence of the part of the unlabeled images. In this paper, the classical method is extended and optimized, and the hierarchical structure construction and update are realized when some object labels are unknown. The convolutional neural network (CNN) is used to encode these images, and the semi-supervised learning method is proposed. The hierarchical structure of the image set which has known the object labels is constructed according to the traditional algorithm. Through the periodic similarity comparison, the unlabeled images in the hierarchy are clustered. The construction of the semi-supervised layer-wise model (SLM) is realized. This paper adopts the real public data sets. The experimental results show that the SLM can effectively realize the construction and update of the hierarchical structure, and can achieve good prediction classification effect on the smaller scale data sets.

Abstract:Image feature is the key to content-based image retrieval (CBIR). Most of the used manual features are difficult to effectively represent the features of the breast mass, and there is a semantic gap between the underlying features and the high-level semantics. In order to improve the retrieval performance of CBIR, this paper uses deep learning to extract the high-level semantic features of images. Because the deep convolution features of mammograms have some redundancies and noises in the spatial and feature dimensions, this paper optimizes the spatial and semantic features of depth features based on the vocabulary tree and inverted files, and constructs two different depth semantic trees. In order to fully exert the discriminative power of deep convolution features, the weight of tree nodes is refined according to the local characteristics of breast image depth features, and two node weighting methods are proposed to obtain better retrieval results. In this paper, 2 200 regions of interest (ROIs) are extracted from the digital database for screening mammography (DDSM) as datasets. The experimental results show that the proposed method can effectively improve the retrieval accuracy and the classification accuracy of the mass region of interest, and has good scalability.

Abstract:Effective fusion of medical data from multiple autism research centers contributes to the diagnosis of autism spectrum disorder (ASD), as large multi-site datasets increase the sample size, which facilitates the investigation of the pathological process of ASD. However, the existing methods generally ignore the heterogeneity (i.e., caused by subject populations and different scanning parameters) among diverse data sites, which degrades the effectiveness of model in ASD diagnosis based on multi-site datasets. To address this issue, we propose a novel domain adaptation method for ASD diagnosis based on joint distribution optimal transport (JDOT). Specifically, we alternately treat one site as target domain, and the rest are sources. Afterwards, we perform alignment in source-target domain by seeking a probabilistic coupling between joint feature and label distributions using optimal transport, which is optimized by an alternative minimization approach. Experimental results demonstrate the effectiveness of our method in ASD diagnosis based on multi-site resting-state functional magnetic resonance imaging (rs-fMRI) datasets.

Abstract:In multi-label learning, feature selection is an effective method to deal with high-dimensional data problems and improve classification performance. However, most of the existing feature selection algorithms are based on the assumption that the label distribution is roughly balanced, and rarely consider the problem of unbalanced label distribution. To solve this problem, this paper proposes a multi-label feature selection algorithm with weakening marginal labels (WML). The algorithm calculates the frequency ratio of positive and negative labels under different labels as the weight of the label, weakens the marginal label by weighting method, and integrates the label space information into the process of feature selection to obtain a more efficient feature sequence, thus improving the accuracy of label description of samples. The experimental results on several datasets show that the proposed algorithm has certain advantages. The effectiveness and rationality of the proposed algorithm are further proved by stability analysis and statistical hypothesis test.

Abstract:As a typical classification method， support vector machine (SVM) has been widely used in various fields. However， the standard SVM faces the following problems in the classification decision: First， it does not consider the distribution characteristics of the classification data; Second， it ignores the relative relationship between sample categories； Third， it can not solve the problem of large-scale classification. In view of this， the rank preservation learning machine based on data distribution fusion (RPLM-DDF) is proposed， in which within-class scatter is introduced to describe the distribution properties， and through the relatively constant position of all kinds of sample data centers， the global sample order remains unchanged. The large-scale classification problem is solved by certifying RPLM-DDF and the duality of the core vector machine. The comparison experiments on the artificial datasets， small-scale datasets and large-scale datasets verity the effectiveness of the RPLM-DDF.

Abstract:A novel extension of residual learning is presented for deep networks which effectively improves the robustness of the learned representation. The method integrates a plug-and-play module， that is, a grouped convolutional encoder-decoder， as additional shortcuts to the original residual architecture. Due to the down-sampling in encoder stage， the decoder modules are driven to produce focally activated feature maps， which highlights the most discriminative regions of input images， and imposes local enhancement on input features through element-wise addition. For efficient model design， we exploit lightweight counterparts by removing part channels of residual mappings， without showing obvious performance degradation. We obtain consistent accuracy gain for various residual architectures with comparable or even lower model complexity.

Abstract:Relation extraction is an important research in the natural language processing (NLP) area. The constituency grammar information, which is widely believed by the academic community, has an important influence on relation extraction. However, there is no obvious effect when the phrase syntactic tree is applied to the relation extraction task. There are two main reasons for this: First, the generalization ability of the constituency parser is poor, which will cause error propagation and then affect its effectiveness in the relation extraction; Second, there are flaws in the way of the use of the phrase syntactic features in the relation extraction task,that is the phrase syntactic structure information learned by the constituency parser is lost, or the wrong influence on the relation extraction is increased. This paper proposes a Chinese relation extraction method based on constituency vector representation to solve the above two problems. The method embeds the text representation learned by the constituency parser into the relation extraction model, thereby improving the relation extraction performance. This paper validates the method on a public Chinese relation extraction data set.

Abstract:Quality Phrase mining is a process of extracting meaningful phrases from text corpus, which is the basis of tasks such as document summary and information retrieval. However, the existing unsupervised phrase mining methods have problems of low quality of candidate phrases and average distribution of feature weight of Quality Phrase. Therefore, a Quality Phrase mining method based on statistic features is proposed. This method combines frequent N-Gram mining, combinatorial constraints of multi-word phrases, and spell checking to ensure the quality of candidate phrases. The public knowledge base is introduced to add labels to the candidate phrases, and the weight distribution of Quality Phrase is realized. The penalty factor is set to adjust the weight ratio considering the mutual influence between the features. The Quality Phrase is extracted according to the score of the feature weighting function of the candidate phrases. Experimental results show that the Quality Phrase mining method based on statistic features significantly improves the precision of phrase mining. Compared with the optimal unsupervised phrase mining methods, the precision, recall and F1-Score values are improved by 5.97%, 1.77%, and 4.02%, respectively.

Abstract:Several complex networks are usually designed in salient region detection to detect saliency, which inevitably leads to very high computational and storage costs. The deep learning network has the characteristics of multi-scale and different convolution layers have different spatial resolutions，thus the design of complex network structure can be avoided. In this paper, a novel convolution neural network is designed by taking advantage of multi-scale characteristics. Both the multi-scale features and the influence of the size of salient regions are considered to saliency detection. Experiments show the superiority of our method on popular benchmark datasets.

Abstract:Traditional collaborative filtering algorithms suffer from data sparsity and cold start problems. Taking advantage of rich information in social networks brings an opportunity to alleviate the problems of data sparsity and cold start. However, the traditional social network-based collaborative filtering algorithm only use the coarse-grained and sparse trust relationships to improve recommendation quality, i.e. they only utilize 0 or 1 to denote the trust relationships between users. In addition, the traditional social network based recommendation algorithms only integrate explicit trust relationships, and ignore implicit trust relationships. In this paper, we propose a graph embedding model based collaborative filtering algorithm. Specifically, we adopt the graph embedding technique to learn the low-dimensional embedded representations of users in social networks, and infer the fine-grained trust relationship between users based on the low-dimensional embedded representations. Finally, the user’s rating of the target item is predicted based on the scoring weights of the target item by the trusted user and the similar one. Experimental results on the actual data sets prove that the performance of the collaborative filtering algorithm based on the graph embedding model is better than that of the traditional collaborative filtering algorithms.

Abstract:The traditional wood defect location methods mainly include physical equipment detection and traditional computer technology detection, but they are difficult to collect data, highly dependent on the data itself, which are not suitable for actual production. We propose an automatic defect location model (ADLM) based on deep learning, which includes single defect location model (SDLM) and multi-defect location model (MDLM) to meet different requirements. This model uses MobileNet as the backbone network, and only a few data sets are needed for training. In the public data set Wood Defect Database, this model has a defect identification rate of 86.1%.In the single defect data set, positioning accuracy of the model can achieve 97.5%. In the multi-defect data set, positioning accuracy of the model can achieve 90.0%. Compared with the traditional model, the ADLM need not manual feature extraction at the early stage, and has the advantages of faster detection speed, higher accuracy and wider applicability.

Abstract:To improve the accuracy of edge detection and ensure the efficiency and effect of image segmentation, an ant colony image enhancement algorithm based on fuzzy clustering is proposed on the basis of ant colony algorithm. The algorithm uses component grayscale value, grayscale gradient value and domain eigenvalue to extract image features, then uses fuzzy clustering to specify the clustering center to improve the convergence speed, then uses the ant colony algorithm to realize the image edge detection, in the process of detection, using the path selection strategy to search the ant colony in order to improve the search efficiency, According to the pheromone update strategy, the optimal path information exchange is realized in order to achieve the purpose of edge point extraction and retrieval, and finally the processed grayscale edge graph coincides with the original picture to realize the effect of image enhancement. Experimental results show that the improved algorithm improves the retrieval time compared with the traditional ant colony algorithm by 20.7%, improves the accuracy by 14.8%, and the texture is clearer in the aspect of image segmentation.

Abstract:Although great progress has been made in the field of computer vision target tracking, the performance of out-of-plane rotation and shape change in video tracking need to be improved. Here, HOG feature based on directional gradient histogram is proposed. Combined with the gray value of the image, the HOG feature is fused and decomposed to improve the performance of the deformation and scale transformation of the video tracking. Firstly, the 31-dimensional features of the HOG and the gray value of the image are extracted from the target region. Secondly, the gray value is regarded as one-dimensional feature, then the gray value is fused with HOG feature into 32-dimensional vector HOG32. Then the HOG32 is decomposed into two parts, namely, HOG1 and HOG2. Finally, compared the response values of HOG1,HOG2 and HOG32, the maximum position is selected as the position of the next frame predicted. The experiment is compared with the other five algorithms on OTB-2013 and OTB-2015 datasets. The results demonstrate that our method achieve better results in out-of-plane rotation, deformation and complex background.

Abstract:Aiming at the problem that the Generalized Hough transform (GHT) is difficult to adaptively terminate in multi-target detection, an adaptive termination algorithm for GHT multi-target detection based on the local peak rate of change in Hough space is proposed. The algorithm is mainly based on the rule that the difference between the local peaks in the target region in the Hough space of the image for detecting is small, and the difference between the local peaks in the target region and the non-target region is big, which leads to the GHT algorithm terminating adaptively without setting a threshold. And the main steps of the algorithm are as follows: Firstly, the cumulative matching distribution of the target in the original image is obtained by the GHT algorithm. Then the distribution results are sorted in descending order. After the sorted distribution is obtained, multiple target recognition results are adaptively detected according to the average change rate of the accumulated peaks, and the algorithm is terminated. Experiments show that compared with the traditional algorithm, this algorithm can accurately detect the multi-target information of the image without significantly increasing the complexity of the algorithm, and can realize the adaptive termination of the multi-target detection algorithm.

Abstract:Aiming at the influence of excessive redundant and unrelated attributes on the diagnosis of lung tumors and the fact that Pawlak rough set is only suitable for dealing with discrete variables and causing a large loss of original information, a high-dimensionality of lung tumors with mixed information gain and neighborhood rough set is proposed.The algorithm first extracts the 104-dimensional feature structure decision information table of 3 000 CT images of lung tumors. With the information gain result, the high correlation feature subset is selected, and the high redundancy attribute is eliminated by the neighborhood rough set. The optimal feature subset is obtained through two attribute reductions. Finally, the support vector machine optimized by the grid optimization algorithm is used to construct the classification recognition model to identify the benign and malignant lung tumors.The feasibility and effectiveness of the method are verified from the two aspects of reduction and classification, and compared with the non-reduction algorithm, Pawlak rough set, information gain and neighborhood rough set reduction algorithm.The results show that the accuracy of the hybrid algorithm is better than other comparison algorithms, the accuracy is 96.17%, and the time complexity is effectively reduced. It has certain reference value for computer-aided diagnosis of lung tumors.

Abstract:To better study the non-independent and identically distributed multi-scale categorical data sets, based on the unsupervised coupling measure similarity method, a multi-scale clustering mining algorithm for non-independent and identically distributed classification attribute data sets is proposed. Firstly, the data set of benchmark scale is clustered based on coupled metric similarity method. Secondly, scale conversion algorithms upscaling based on single chain and downscaling based on Lanczos kernel are proposed for scale conversion. Finally, experiments are performed using the public data sets and the real data sets of the H province. In the experiment, couple metric similarity (CMS), inverse occurrence frequency (IOF), hamming distance (HM) and other similarity metric methods combined with spectral clustering algorithm are compared and the experimental results demonstrate that the NMI value of the upscaling increases by 13.1%, the mean of MSE value reduces by 0.827, and the mean of F-score value increases by 12.8%. Compared with other comparison algorithms, the mean of NMI value of downscaling increases by 19.2%, the mean of MSE value reduces by 0.028, and the mean of F-score value increases by 15.5%. Experimental results and theoretical analysis show that the proposed algorithm is effective and feasible.

Abstract:At present, bird activity failure has become one of the main hidden dangers of high-speed railway. Finding and cleaning the birds’ nest of the catenary is a countermeasure. Traditional birds’ nest object detection methods require manual extraction of features, but hand-designed features are difficult to ensure generalization in complex contact network scenarios. To solve this problem, this paper proposes to use the deep learning based object detection algorithm to identify the birds’ nest on catenary. At the same time, an improved model based on the one-stage object detection model RetinaNet is proposed. The P2 feature layer is added to expand the receptive field range of the network, so that the smaller nest can be better detected. Finally, these deep learning based object detection algorithms are trained and tested using data sets collected by on-board equipment of high-speed railways. Experimental results show that the object detection algorithm based on deep learning is excellent in the catenary birds’ nest detection task, and the improved RetinaNet model has a mAP value of 90.4%, which is better than the original model. This algorithm has certain both reference and application value for the obstacle avoidance task of high?-?speed railway.

Abstract:Lithology identification is a key and difficult problem in reservoir geological interpretation. The development and application of artificial intelligence, especially machine learning technology, provides a new technical way to solve lithology identification problems. This paper uses machine learning models such as support vector machine (SVM), multi-grained cascade forest (GCForest), random forest (RF) and eXtreme gradient boosting (XGBoost) to build a heterogeneous multi-layer integrated learning model. The integrated learning model overcomes the shortcomings of single model such as high requirement for data sets, poor generalization ability and low recognition accuracy. In this paper, lithology recognition experiments are carried out using integrated models and single models. The experimental results show that the average accuracy of the integrated model is 96.66%, higher than that of SVM (75.53%), GCForest (96.21%), random forest (95.06%) and XGBoost (95.77%). The integrated model can be effectively applied to lithology identification and classification tasks in reservoir geological analysis with strong adaptability and high recognition accuracy.

Abstract:Recognizing predicate verbs is the key to understanding sentences. Because Chinese predicate verbs are complex in structure, flexible in use, and changeable in form, identifying predicate verbs is a challenging task in Chinese natural language processing. This article introduces the concepts related to the recognition of Chinese predicate verbs from the perspective of information extraction, and proposes a method for marking Chinese predicate verbs. On this basis, a Chinese predicate verb recognition method based on Attentional-BiLSTM-CRF neural network is studied. This method uses the bidirectional recurrent neural network to obtain the dependency relationship within the sentence, and then uses the attention mechanism to model the focus role of the sentence. Finally, a maximized labeling path through the conditional random field(CRF)layer is returned. In addition, in order to solve the problem of the uniqueness of predicate verb output, a unique recognition model of predicate verb based on convolutional neural network is proposed. Through experiments, the algorithm exceeds the traditional sequence labeling model CRF, and reaches an F value of 76.75% on the Chinese predicate verb data labeled in this paper.