Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2020-09-28 Cooperative journals: 《计算机应用研究》
Abstract： Person re-identification is to judge whether the pedestrian across different cameras belongs to the same person or not. While it is challenging task due to the large variations in person pose, occlusion, background clutter, etc. And several deep learning based person re-identification have been proposed and achieved remarkable performance. However, these methods are only considered separately from the local or global features of the pedestrian, ignoring the relationship between the features. So this paper proposed the enhanced feature convergent network (EFCN) . In the global branch, the paper used to employ the new attention to pay close attention to the global feature of pedestrians. In the local branch, it proposed the gated recurrent unit change network(GRU-CN) to obtain more robust local features, and then this paper used feature fusion to connect the extracted global and local features. Extensive comparative experiments show that EFCN can achieve better experimental results on three standard person Re-ID datasets. The proposed enhanced feature convergent network can extract highly discriminative pedestrian features. This model can be applied to the problem of Re-ID under non-overlapping multi-cameras in large scenes. It has high recognition ability and accuracy. The method can extract robust features for pedestrian images with changing background.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2020-09-28 Cooperative journals: 《计算机应用研究》
Abstract： This paper developed a causality modeling algorithm based on maximum information transfer entropy to solve the problem that traditional causality algorithms were difficult to accurately analyze non-linear data with a lot of noise. First, used the maximum information coefficient to detect the correlation between time series trends of non-linear data. Weaken the effect of noise on the correlation between variables. Secondly, eliminated weakly related variables based on screening factors. Calculated the transfer entropy between strong correlations using stochastic empirical valuation. Thereby reducing the calculation amount of transfer entropy. Finally, transfer entropy determined causal direction. Formed a one-way causal network that supports link traceability. Test analysis of the algorithm using classic chemical process data sets. Test results show that, compared to existing algorithms, this algorithm can locate abnormal variables. The stability of this algorithm for modeling high-dimensional data of more than 12 dimensions is higher than 85%, and the accuracy rate of causality can reach 83.33%. The actual modeling effect of this algorithm is better than the comparison algorithms, and it can detect and locate industrial control system abnormalities.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-04-01 Cooperative journals: 《计算机应用研究》
Abstract： To solve the problem that the existing step counting algorithms have poor adaptability to different motion states, this paper proposed a step counting algorithm based on self-adaptive threshold. By using the 3-axis accelerometer sensor in smart bracelet, the proposed algorithm first performed the acceleration data collection when users walked at three frequencies including slow walking, fast walking and running. After five-point filter preprocessing, this algorithm detected peaks and valleys in the self-adaptive time window. Then it taken the average of the peak mean and valley mean as the upper threshold, and used the valley mean as the lower threshold. Based on this, the algorithm adopted a dynamic threshold analysis to achieve the step counting. Finally, in this algorithm, false steps could also be detected according to the regularity of walking amplitude and frequency. The test shows that the achievable average step counting accuracy of the proposed algorithm is above 91.88% for different users at three different walking frequencies.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-07-23 Cooperative journals: 《计算机应用研究》
Abstract： Existing research on the reliability allocation of complex software system just based on software systems whose structure were fixed, but the structure of software systems always change in practice. For this contradiction, this paper presents a dynamic reliability allocation optimization model of complex software system. At the same time, developing a dynamic reliability allocation algorithm of complex software system based on differential evolution. Specifically, when the system structure changes, we first estimate the parameters of each module according to the Dempster-Shafer evidence theory. Next, taking account of the relevance of the system before and after the change, we retain some historical solutions in the previous search to form part of the initial population in differential evolution. Finally, verifying the effectiveness of the proposed approach by simulation experiments.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-05-20 Cooperative journals: 《计算机应用研究》
Abstract： In the traditional Web page ranking algorithm Okapi BM25, there exists a problem that the retrieval results are independent to the domain keywords, and the improved algorithm needs to build the domain vector manually. To address this issue, we propose a web page ranking algorithm based on BM25 and softmax regression classification model. In this method, we first encode the web page text with the bag-of-words model, and then train the softmax regression classification model by a small amount of web data to predict the category scores of the test web data. Finally we combine the category scores and the BM25 information retrieval scores to get the final ranking of web page results. Experiment results show that our method can meet the user's information need better without even manually creating the domain vector.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-05-20 Cooperative journals: 《计算机应用研究》
Abstract： Multi-label learning has been widely used in many application scenarios right now. In this kind of learning problem, each instance is simultaneously assigned with more than one class label. Since different class labels might have their own unique characteristics (i. e. , label-specific feature) which would be more useful for label classification, so some multi-label learning approaches based on label-specific features had already been proposed. Therefore, aiming at the problem that redundant feature space caused by label-specific feature construction, a multi-label label-specific feature extraction algorithm named LIFT_RSM is proposed, which can improve the performance of classification by comprehensively using random subspace method and the thought of pair-wise constraint dimensionality reduction to extract effective feature information in label-specific feature space. The experimental results on several datasets show that the proposed algorithm can achieve better classification results compared with several classical multi-label algorithms.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-05-02 Cooperative journals: 《计算机应用研究》
Abstract： Aiming at the excessive location error caused by heterogeneous terminal (fingerprint database terminal and test terminal) in WLAN indoor localization system, a solution based on DBSCAN-GRNN-LSSVR algorithm is proposed. In this paper, the least square support vector regression (LSSVR) algorithm is employed to build the mapping relationship model between the received signal strength (RSS) of fingerprint database terminal and physical coordinate locations. The scatter plot is obtained through listing the RSS values collected by the fingerprint database terminal and the test terminal at the calibration point. The boundary points and noise points are eliminated by density-based spatial clustering. Generalized regression neural network (GRNN) is used to construct the heterogeneous terminal mapping function of the RSS. The LSSVR model is used to determine the location of the test point. Proved by experiment, compared with LSSVR algorithm, using the proposed DBSCAN-GRNN-LSSVR algorithm to calibrate the heterogeneous terminal, test terminal positioning accuracy increased by 18-40%, which effectively solves the problem of excessive localization deviation caused by heterogeneous terminals.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-04-17 Cooperative journals: 《计算机应用研究》
Abstract： In view of the fact that the existing sentence vector learning method can not well learn the relational knowledge information and express the complicated semantic relation, this paper proposed a relational information sentence vector model (RISV) based on the PV-DM model and the relational information model. This model used the PV-DM model as the basic model of sentence vector training, and then added the knowledge constraint of relational information to make the improved model can learn the relationship between the words in the text and uses the RCM model as Pre-training model to further integrate the information of the semantic relationship constraints, and finally validates the validity of the RISV model in two tasks: document classification and short text semantic similarity. The experimental results show that sentence vectors learned by RISV model can better represent the text.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-04-17 Cooperative journals: 《计算机应用研究》
Abstract： To deal with the problem that classical spectral clustering algorithms are unable to determine the number of clusters automatically, and low efficiency in processing large amount of data with. This paper proposes a spectral clustering algorithm based on the optimization of density peak value. The method firstly calculates the local density of data object and the minimum distance between each data object and other data objects. Adaptive clustering algorithm is generated to determine the number of clusters and to optimize the number of clusters according to certain rules. Secondly, adopting Nystr鰉 sampling can reduce the time complexity of characteristic decomposition and improve the efficiency of the algorithm. The experimental results show that this method can accurately obtain the number of clusters and effectively improve the accuracy and efficiency of clustering effectively.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-04-12 Cooperative journals: 《计算机应用研究》
Abstract： The existing AdaBoost. M2 algorithm are insufficient in protocol traffic multiclass imbalance to solve the problem. So, this thesis proposes an ensemble algorithom called RBWS-ADAM2 for the classification of multiclass internet traffic. During each iteration of AdaBoost. M2, this algorithm preprocessed the training dataset by randomly balanced resampling, this strategy changed the number of majorities and minorities by randomly setting the sampling balance point to build multiple different training datasets. Moreover, this strategy toke sample weight as the basis for sample screening to strengthen the learning of this kind of sample. The experimental comparison of RBWS-ADAM2 algorithm and other similar algorithms on the internationally published protocol traffic datasets shows that, compared to other algorithms, the proposed RBWS-ADAM2 algorithm not only improves the F-Measure of most minorities, but increases the overall G-mean and the overall average F-measure effectively, and obviously enhances the overall performance of the ensemble classifier.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-04-12 Cooperative journals: 《计算机应用研究》
Abstract： The spatial and angular resolution are comparatively low owing to the limitation of the resolution of its pixel sensor which is due to design reasons. To address the above problem, this paper presented a new super-resolution method based on global and local views to simultaneously enhance both the spatial and angular resolutions of a light field image. First of all, it adaptively chose local views according to the position of the novel view, and then used spatial super-resolution convolution neural network to get high resolution sub-images such as global-view and local-view images, subsequently it extracted and fused the depth and color features from the image obtained by backward warping the input views. Then put the results into the angular super-resolution network which produced a novel view image as the final result. The experimental results demonstrate that the reconstruction effect of this method is superior to the state-of-the-art’s. And in peak signal to noise ratio get 3dB promotion and structural similarity image measure upgrade about 0.02. In addition, it is not only able to improve spatial and angular resolution of multi-view images, but also to maintain edge information of the novel view, thus effectively avoiding the loss of objects and yielding a better reconstruction effect.
Hits
Hits
Downloads
Comment