Loading...

Table of Content

      
    20 April 2019
    Volume 49 Issue 2
    Machine Learning & Data Mining
    Fast 4-points congruent sets for coarse registration of 3D point cloud
    Shiguang LIU,Hairong WANG,Jin LIU
    Journal of Shandong University(Engineering Science). 2019, 49(2):  1-7.  doi:10.6040/j.issn.1672-3961.0.2018.244
    Abstract ( 2244 )   HTML ( 146 )   PDF (6934KB) ( 711 )   Save
    Figures and Tables | References | Related Articles | Metrics

    In order to solve the problem that the 4-points congruent sets (4 PCS) method suffered from low computational efficiency and high registration errors when the overlap rate of two pieces of input point clouds was low, fast 4-points congruent sets (F-4PCS) was put forward. A new method for selecting four-point basis was presented. The source point cloud and target point cloud were given, their boundaries were separately extracted and extended as the boundary feature bands, and then a consistent four-point basis set was chosen from the boundary feature bands. This method could avoid some unnecessary iterations. By limiting the characteristics of the four-point basis, the invalid four-point basis was removed, it could reduce the verification time of the algorithm and improve the computational efficiency. Experiments results carried out on the relevant data sets showed that the F-4PCS method was more efficient than conventional 4PCS method in the case of low overlap rate of input point clouds and the registration success rate was higher than state-of-the-arts.

    Image attribute annotation based on extreme gradient boosting algorithm
    Hongbin ZHANG,Diedie QIU,Renzhong WU,Tao ZHU,Jin HUA,Donghong JI
    Journal of Shandong University(Engineering Science). 2019, 49(2):  8-16.  doi:10.6040/j.issn.1672-3961.0.2018.271
    Abstract ( 860 )   HTML ( 7 )   PDF (4657KB) ( 324 )   Save
    Figures and Tables | References | Related Articles | Metrics

    To improve annotation performance, a novel image attribute annotation model based on eXtreme gradient boosting (XGBoost) algorithm was proposed: image features i.e. local binary patterns (LBP), Gist, scale invariant feature transform (SIFT), and visual geometry group (VGG) were extracted respectively to better characterize the key visual content of images. Then the state-of-the-art boosting algorithm called XGBoost was used to design a strong classifier by integrating a group of weaker classifiers. Based on the strong classifier, image attribute annotation was implemented. A lot of valuable deep semantic implied by image attribute was mined in turn to create a novel hierarchical attribute representation mechanism, which was closer to human's objective cognition. Finally, transfer learning strategy was designed to further improve annotation performance. Experimental results showed that the key visual content of images was truly characterized by the Gist feature. Compared to the best competitor before transfer learning, the accuracy of basic transfer (BT) learning strategy was improved about 8.69%. Compared to the best competitor of BT, the accuracy of hybrid transfer (HT) learning strategy was improved about 17.55%. The annotation accuracy was improved by the presented model.

    The vulnerability mining method for KWP2000 protocol based on deep learning and fuzzing
    Chengbin ZHANG,Hui ZHAO,Zongyu CAO
    Journal of Shandong University(Engineering Science). 2019, 49(2):  17-22.  doi:10.6040/j.issn.1672-3961.0.2018.340
    Abstract ( 900 )   HTML ( 17 )   PDF (1273KB) ( 325 )   Save
    Figures and Tables | References | Related Articles | Metrics

    A kind of vehicle-onboard diagnosis Protocol standard, keyword protocol 2000 (KWP2000) KWP2000, was investigated in details. KWP2000 was widely used in the automobile industry and the loophole of possible communication Protocol. We analyzed the current situations of the fuzzing, and based on this, we proposed a generative adversarial networks (GAN) by deep learning neural network for automobile body network KWP2000 protocol hole mining method. The forward feedback network was closeted as the generation model, and the support vector machine was used as the discriminant model. We used the neural network model to train the test case data of the KWP2000 protocol data, the fuzzing of KWP2000 was carried out by using these test case data. Through experiments, we found that the target protocol KWP2000 had long loopholes, coding errors and other vulnerabilities. Experimental results showed that this fuzzing method was efficient and safe.

    Survey of human-robot interaction control for autonomous driving
    Qijie ZOU,Haoyu LI,Rubo ZHANG,Tengda PEI,Yan LIU
    Journal of Shandong University(Engineering Science). 2019, 49(2):  23-33.  doi:10.6040/j.issn.1672-3961.0.2017.503
    Abstract ( 983 )   HTML ( 10 )   PDF (1268KB) ( 619 )   Save
    Figures and Tables | References | Related Articles | Metrics

    This article summarizes the machine learning methods of human-robot interaction in autonomy vehicles. By introducing the value and significance of human-robot interaction, the relationship between the human-robot interaction problem definition and machine learning were identified, the human-robot interactions team framework was built. The frameworks of human-robot interaction and the research methods of autonomous driving system were reviewed, the general structure for solving human-robot interaction problems was presented. Furthermore, its machine learning algorithm from the two aspects of autonomous control system and driver modeling was introduced. The prospects of the future research direction were summarized.

    Chinese short text classification method based on word2vec embedding
    Mingxia GAO,Jingwei LI
    Journal of Shandong University(Engineering Science). 2019, 49(2):  34-41.  doi:10.6040/j.issn.1672-3961.0.2018.197
    Abstract ( 2119 )   HTML ( 49 )   PDF (2868KB) ( 307 )   Save
    Figures and Tables | References | Related Articles | Metrics

    In the short text classification process, the weak feature expression of the limitation of the number of words restricted the classification effect. To solve this problem, a Chinese short text classification method based on embedding trained by word2vec from Wikipedia (CSTC-EWW) was proposed, and a series of experiments for short texts with 4 topics from the iask.com website were finished. This method firstly trained the embedding by word2vec from Wikipedia corpus. the feature of short text based on the embedding was established. Naive Bayes and SVM was used to classify short text. The experimental results showed the following conclusions: CSTC-EWW could effectively classify short texts and the best F-value could reach 81.8%; Comparing the text feature expression of BOW model weighted by TF-IDF and the method of extending feature from Wikipedia, the classification results of CSTC-EWW were significantly better and F-measure of CSTC-EWW on car could be increased by 45.2%.

    Recommendation algorithm based on trust network reconfiguration
    Yun HU,Shu ZHANG,Hui LI,Kankan SHE,Jun SHI
    Journal of Shandong University(Engineering Science). 2019, 49(2):  42-46.  doi:10.6040/j.issn.1672-3961.0.2018.346
    Abstract ( 630 )   HTML ( 6 )   PDF (1715KB) ( 294 )   Save
    Figures and Tables | References | Related Articles | Metrics

    A new recommendation algorithm was investigated base on the problem of trust network reconfiguration. The initial trust network was constructed by combining the user similarity value with the trust relationship, and the initial prediction of the user's unrated items was carried out.A method based on reliability was used to evaluate the quality of prediction score. The unrated items were predicted according to the new user trust network. The performance was verified on two real data sets, which were Epinions dataset and Flixster dataset. The experimental results showed that the reconfiguration algorithm of trust network effectively solved the problem of data sparsity in recommendation system, and it was superior to the traditional recommendation algorithm in recall and precision ratio.

    Epsilon truncation algorithm based on NDX and adaptive mutation operator
    Jin LI,Erchao LI
    Journal of Shandong University(Engineering Science). 2019, 49(2):  47-53.  doi:10.6040/j.issn.1672-3961.0.2018.194
    Abstract ( 604 )   HTML ( 9 )   PDF (1260KB) ( 376 )   Save
    Figures and Tables | References | Related Articles | Metrics

    It was hard for constrained optimization algorithm to maintain a good balance of convergence and distribution. To solve this problem, a ε-truncation algorithm based on the normal distribution crossover (NDX) and adaptive mutation operator was proposed, which introduced the normal distribution into the simulated binary crossover (SBX) operator, so that the algorithm could search wider space and easily jump out of local optimum. The proposed algorithm combined the adaptive mutation operator with the current information of the individual in the population, and guided the population evolving toward the real Pareto front. Moreover, it preserved the Pareto optimal solutions and a certain number of infeasible solutions with the adaptive ε truncation strategy. At the same time using the information of these infeasible solutions, it increased the search intensity of space and improved the diversity of population. According to three standard test functions experimental results, the solution set in this study could well track the real Pareto solution set. The proposed method could effectively coordinate the convergence and distribution of the algorithm.

    Facial age estimation based on multivariate multiple regression
    Run XIANG,Sufen CHEN,Xueqiang ZENG
    Journal of Shandong University(Engineering Science). 2019, 49(2):  54-60.  doi:10.6040/j.issn.1672-3961.0.2017.420
    Abstract ( 693 )   HTML ( 12 )   PDF (1956KB) ( 225 )   Save
    Figures and Tables | References | Related Articles | Metrics

    Label distribution learning based facial age estimation model was an effective method to solve the problem of insufficient training data caused by the difficulty of facial image collection, where its motivation was that facial aging information on adjacent ages can be introduced to enhance the age estimation model due to human faces changing slowly. Given a certain age to learn, label distribution learning converted the learning target from a continuous value to an age label distribution vector, which was generated according to the description degree of the neighboring ages. However, the existed methods had the drawbacks of separated age prediction model (maximum entropy based methods) or tending to be overfitting (neural network based methods). So a method of facial age estimation based on multivariate multiple regression was proposed, the label distribution learning based age estimation problem was transformed into a multivariate multiple regression analysis task and then solved by the multivariate partial least squares regression. Multivariate partial least squares regression had no assumption about the data distribution and built an integrated effective model for all ages even when there is a strong correlation among independent variables. Extensive comparative experimental results on FG-NET facial age estimation dataset showed that the proposed method significantly improved the training efficiency, and at the same time, had higher age estimation accuracy than the state-of-the-art methods.

    Images auto-encoding algorithm based on deep convolution neural network
    Yijiang HE,Junping DU,Feifei KOU,Meiyu LIANG,Wei WANG,Ang LUO
    Journal of Shandong University(Engineering Science). 2019, 49(2):  61-66.  doi:10.6040/j.issn.1672-3961.0.2017.432
    Abstract ( 1556 )   HTML ( 35 )   PDF (1646KB) ( 505 )   Save
    Figures and Tables | References | Related Articles | Metrics

    At present, image coding research was focused on information lossless, but it did not reflect the social network image differentiation. A novel social network images auto-encoding algorithm based on deep convolution neural network was proposed. The algorithm obtained good performance on image auto-encoding, which combined the feature extraction ability of deep convolutional neural network and characteristics of images in social networks. It combined the characteristics of the social network image with the clustering algorithm to cluster social network image and got the distance information, next the deep convolutional neural network was used to learn the distance information of these images, then it extracted the fully connected layer in the deep convolution neural network as the image coding, repeated the above steps and got the image coding finally. The experimental results showed that the proposed algorithm performed better than other algorithms of image search, and was more adaptive in the social network image search than that of the other algorithms mentioned.

    Automatic landmarks identification and tracking of bat flight
    Xu YANG,Hui CHEN,Yousi LIN,Changhe TU
    Journal of Shandong University(Engineering Science). 2019, 49(2):  67-73.  doi:10.6040/j.issn.1672-3961.0.2018.155
    Abstract ( 713 )   HTML ( 10 )   PDF (2680KB) ( 348 )   Save
    Figures and Tables | References | Related Articles | Metrics

    Bats could serve as an inspiration for flapping-wing air vehicles. Understanding bats flight with computer vision techniques required a large copious of fiducial landmarks. Thus, accuracy of landmark identification and tracking was critical to bat flight research. General low-level feature extraction methods based on local extrema often resulted in high false positives. A landmark identification method based on image segmentation was proposed. An initial bat silhouette was first obtained using frame difference and then refined by compensating camouflage parts. The landmarks were enhanced by LoG operation. Finally, the coordinates of landmarks were computed from the centroids of connected components. Furthermore, a landmark tracking method based on ICP (Iterative Closet Points) was proposed. Bat region was divided into several parts, in which landmarks were aligned by ICP. The correspondences were determined by the nearest neighbor search. The method reached an identification accuracy up to 96%, and could track the landmark correctly when occlusion wasn′ occurred, which was better than SIFT, BRISK, and optical flow tracking methods.

    Computer aided diagnosis method for breast cancer based on AlexNet and ensemble classifiers
    Xiaoxiong HOU,Xinzheng XU,Jiong ZHU,Yanyan GUO
    Journal of Shandong University(Engineering Science). 2019, 49(2):  74-79.  doi:10.6040/j.issn.1672-3961.0.2018.273
    Abstract ( 658 )   HTML ( 12 )   PDF (1350KB) ( 226 )   Save
    Figures and Tables | References | Related Articles | Metrics

    In order to solve the manual feature extraction of medical images in computer aided diagnosis, Alexnet was pre-trained on the ImageNet dataset, and feature extraction was performed on the medical image based on Alexnet with transfer learning. The ensemble learning method was used to train the classifier to classify and obtain a better classification effect than the single classifier. The results showed that the AUC(area under curve) of Alexnet deep learning model and random forest ensemble classifier reached 0.87±0.03, and the effect of the ensemble classifier was better than that of the single classifier in the same network depth.

    Research onfeature selection technology in bearing fault diagnosis
    Jiachen WANG,Xianghong TANG,Jianguang LU
    Journal of Shandong University(Engineering Science). 2019, 49(2):  80-87, 95.  doi:10.6040/j.issn.1672-3961.0.2018.268
    Abstract ( 945 )   HTML ( 88 )   PDF (3981KB) ( 268 )   Save
    Figures and Tables | References | Related Articles | Metrics

    A new method based on feature selection (FS) was proposed to select efficient features to promote the classification accuracy in bearing fault diagnosis. First, the outstanding features whose classification accuracy were higher than the threshold were directly selected by diagnosis model from a big feature set. Then the significant combinations of features which had less dimensions and higher classification accuracy were selected in the candidate feature set by a distinctive feature-oriented manner. Experiments showed that the proposed method had advantages in selecting efficient features, reducing the model parameters, decreasing the demand of samples and enhancing the model classification accuracy. As a result, it provided a new idea for feature selection and improved the efficiency of bearing fault diagnosis.

    Real-time traffic prediction based on MGU for large-scale IP backbone networks
    Fang GUO,Lei CHEN,Ziwen YANG
    Journal of Shandong University(Engineering Science). 2019, 49(2):  88-95.  doi:10.6040/j.issn.1672-3961.0.2018.342
    Abstract ( 844 )   HTML ( 11 )   PDF (4768KB) ( 325 )   Save
    Figures and Tables | References | Related Articles | Metrics

    In order to overcome the shortcomings of long short-term memory (LSTM) computing cost, a real-time traffic prediction method based on minimum gated unit (MGU) for large-scale IP backbone networks was proposed. The experimental results showed that compared with the LSTM-based traffic prediction method, the proposed method achieved fairly or even better traffic prediction performance with less model training time, meanwhile it outperformed the most advanced feed forward neural network (FFNN), LSTM and gated recurrent unit(GRU) in terms of prediction accuracy and real-time performance.

    An unsupervised color image segmentation method based on fusion of multiple methods
    Xinyu DONG,Hanyue CHEN,Jiaguo LI,Qingyan MENG,Shihe XING,Liming ZHANG
    Journal of Shandong University(Engineering Science). 2019, 49(2):  96-101.  doi:10.6040/j.issn.1672-3961.0.2018.242
    Abstract ( 686 )   HTML ( 7 )   PDF (6833KB) ( 326 )   Save
    Figures and Tables | References | Related Articles | Metrics

    An unsupervised color image segmentation method based on fusion of multiple methods was proposed, which considered the defects of traditional K-means clustering color image segmentation method, such as the need to set the number of initial segmentation categories artificially and the vulnerability to noise interference, etc. First of all, the original image was processed by spectral information enhancement to improving the efficiency of image information extraction. Next, the number of K-means clustering segmentation categories was determined automatically by using Davies-Bouldin Index, and the clustering analysis was carried out for images and each pixel in an image was labeled. Then, the labeled image was segmented by combining the Gauss-Markov random field theory. Finally, the image after-processing was made based on the morphological operators. The segmentation experiments were carried out by using different methods, the results showed that the segmentation effect of the proposed method was closer to the origin image, and the proposed method had good robustness. And the results of quantitative evaluation of segmentation showed that this method had more advantages in segmentation precision and accuracy.

    A microblog rumor events detection method based on C-GRU
    Lizhao LI,Guoyong CAI,Jiao PAN
    Journal of Shandong University(Engineering Science). 2019, 49(2):  102-106, 115.  doi:10.6040/j.issn.1672-3961.0.2018.189
    Abstract ( 1521 )   HTML ( 39 )   PDF (1249KB) ( 505 )   Save
    Figures and Tables | References | Related Articles | Metrics

    A microblog rumor events detection model based on convolution-gated recurrent unit(C-GRU) was proposed. Combining the advantages of CNN and GRU, the microblog event′s posts was vectorized. By learning the features representation of the microblog windows through the convolution layer of CNN, the features of microblog windows was spliced into a sequence of window feature according to the time order, and the sequence of window feature was put into the GRU to learn feature representation of sequence for rumor events detection. Experimental results from real data sets showed that this model had better ability to rumor detection than other models based on traditional machine learning, CNN or RNN.

    Transfer fuzzy clustering based on self-constraint of multiple medoids
    Jun QIN,Yuanpeng ZHANG,Yizhang JIANG,Wenlong HANG
    Journal of Shandong University(Engineering Science). 2019, 49(2):  107-115.  doi:10.6040/j.issn.1672-3961.0.2018.458
    Abstract ( 605 )   HTML ( 6 )   PDF (1725KB) ( 181 )   Save
    Figures and Tables | References | Related Articles | Metrics

    Transfer clustering approaches derived from the fuzzy C-means (FCM) framework, which considered virtual centers from source domains as transfer knowledge, inherited the shortcomings of FCM. These methods were not robust to outliers and noises, and whose single cluster centers were not sufficient enough to capture the inner structures of clusters. To solve the problems, a transfer fuzzy clustering approach was proposed based on the self-constraint of multiple medoids. Prototype weights were introduced and assigned to each object to capture the inner structures of clusters. Such a weighting strategy could capture the inner structures of clusters more sufficiently and made the clustering more robust to outliers and noises; Furthermore, with the distribution of data in the source domain, the inner structure of data in the target domain was reconstructed, and the corresponding new structure was considered as the transfer knowledge to guide the clustering of the target domain. Relative to the use of single virtual center of each cluster as transfer knowledge, the updated inner structures of data in the target domain contained more knowledge. Experimental results demonstrated that the proposed approach achieved 0.674 5 and 0.608 4 improvements in terms of NMI and ARI on synthetic datasets and real-life datasets compared with introduced benchmarking approaches. Therefore, based on the transfer principle of the self-constraint of multiple medoids, the proposed clustering approach performed well in the transfer environment.

    An error sensitivity model based on video statistical features
    Tong LI,Ran MA,Honghe ZHENG,Ping AN,Xiangyu HU
    Journal of Shandong University(Engineering Science). 2019, 49(2):  116-121.  doi:10.6040/j.issn.1672-3961.0.2018.243
    Abstract ( 692 )   HTML ( 5 )   PDF (2311KB) ( 246 )   Save
    Figures and Tables | References | Related Articles | Metrics

    The traditional packet losses affected the video quality, an error sensitivity model was proposed. For every damaged block, the available statistical features around the block were extracted, which included the losing status of neighboring blocks, texture complexity, motion vector and gradient. After concealing the damaged videos with error concealment methods, error sensitivities were computed. The relationship model between statistical features and error sensitivities was finally established by machine learning technology. Experimental results demonstrated that the proposed model could accurately predict the sensitivities of video frames′ local differences to different packet loss cases, compared with the state-of-art assessment methods, especially for the slow-motion video sequences, the prediction accuracy could be obviously superior to other methods.

    A rule extraction method based on multi-objective co-evolutionarygenetic algorithm
    Zhongwei ZHANG,Hongyan MEI,Jun ZHOU,Huiping JIA
    Journal of Shandong University(Engineering Science). 2019, 49(2):  122-130.  doi:10.6040/j.issn.1672-3961.0.2018.211
    Abstract ( 540 )   HTML ( 8 )   PDF (1309KB) ( 174 )   Save
    Figures and Tables | References | Related Articles | Metrics

    Aiming to deal with the problem that continuous numerical attributes in transaction database were difficult to divide and the efficiency of rule extraction was low, a multi-objective co-evolutionary quantification association rule extraction method was proposed with the cooperative of crossover population and mutation population. The non-dominated sorting of the Pareto principle was used to optimize the individuals of population. Genotype and phenotype of individuals similarity were used to control the matching of individuals in the crossover population. The mutation population was segmented by the concept of level set, then, single point mutation and multiple point mutation were adopted according to the quality of individuals to enhance the individuals diversity. The pareto optimal solution set was obtained from the elite population which was used to preserve the excellent individuals in the crossover population and the mutation population. The simulation results on different datasets showed that the algorithm achieved a good balance of performance and quantity, and the data set was effectively covered, which verifies the effectiveness and feasibility of the algorithm.