Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2022-05-10 Cooperative journals: 《计算机应用研究》
Abstract： To address the problem that traditional similarity-based outlier detection algorithms were not effective enough on high-dimensional unbalanced datasets, this paper proposed a novel Ensemble learning and Random projection-based Outlier Detection (EROD) framework. Firstly, the EROD algorithm integrated several random projection methods to reduce the dimensionality of high-dimensional data, which improved the data diversity. Secondly, it integrated several different traditional outlier detectors to build a heterogeneous ensemble model, which increased the robustness of the algorithm. Finally, the EROD acquired the final outlier value of the object by using the heterogeneous ensemble model to train the reduced-dimensional data and by using two optimal combinations of the trained model to reduce the total error, and the algorithm determined the object with high outlier value as outlier point. The results showed that the algorithm had an average improvement of 3.6% and 14.45% in AUC and Precision@n value compared with the traditional outlier detection algorithm and the outlier detection algorithm based on ensemble learning. Therefore, the EROD algorithm has the advantage of handling the anomalies of high-dimensional unbalanced data.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2022-04-07 Cooperative journals: 《计算机应用研究》
Abstract： Rain streaks can seriously degrade the quality of captured images and affect subsequent computer vision tasks. In order to improve the quality of rainy images, this paper proposed a single-image deraining algorithm based on Transformer. First, the algorithm obtains a wide range of receptive fields through the transformer with window mechanism, and then obtains the contextual information of rain streak features to improve the ability of the model to extract rain streak features; secondly, the algorithm extracts and fuses different kinds and levels of features through multi-branch modules to improve the model's ability to characterize complex rain streaks information; finally, this paper fuses the shallow features and deep features through residual connections to complete the missing details in the deep features, which enhances the expression ability of the network. The experimental results on the public datasets Rain100L, Rain100H and the private dataset Rain3000 show that the method is more effective in removing rain streaks compared to existing algorithms while better recovering the lost background texture information in the images. PSNR and SSIM have respectively reached 38.33/0.9855, 28.42/0.9000 and 34.51/0.9643.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2020-09-28 Cooperative journals: 《计算机应用研究》
Abstract： In order to understand the latest research progress of TransE based representation learning methods in real time, this paper classifies TransE based representation learning methods into four types: the method based on complex relationship, the method based on relationship path, the method based on image information, and the method based on other aspects. Then, this paper analyzes the design ideas, advantages and disadvantages of each method. At the same time, it compares and summarizes the common data sets and evaluation indexes of the TransE based representation learning method, as well as the performance of various TransE based representation learning algorithms in the experiment. Finally, this paper summarizes the research of the whole paper and looks forward to the future research hotspot. From the research results, PaSKoGE method, NTransGH method, TCE method and TransD method perform the best in link prediction and triple classification tasks, which are worth promoting and further expanding, and can be further improved in path specific embedding, two-layer neural network, triple context and dynamic mapping matrix construction.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-01-03 Cooperative journals: 《计算机应用研究》
Abstract： In order to understand the development of single image super-resolution reconstruction (SISR) based on deep learning and grasp the hotspots and directions of the current research, this paper combs the existing model of single image super-resolution reconstruction based on deep learning. Firstly, the paper introduces the related deep learning algorithm, these models based on deep learning and their evaluation index. In addition, it compares the performance of existing models through experiments, which aims to understand the advantages of single-image super-resolution reconstruction model based on deep learning. Finally, the paper summarizes the key issues of single-image super-resolution reconstruction, and prospects the future development trends.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-01-03 Cooperative journals: 《计算机应用研究》
Abstract： In order to understand the latest research progress of the classical clustering algorithm K-means in Spark environment, and grasp the current research hotspots and directions of K-means algorithm, this paper reviews the initial center point optimization research on K-means algorithm. Firstly, it introduces the memory computing framework Spark and K-means algorithms, and analyzes the cause and effects of clustering instability of K-means algorithm, which aimed to point out the importance of optimizing K-means algorithm. And it introduces the main methods and the latest research status of optimizing the initial center point of K-means in Spark environment in detail, and also discusses the future research trends in initial center point optimization of K-means.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-08-13 Cooperative journals: 《计算机应用研究》
Abstract： As an important part of Chinese sign language recognition, the recognition of Chinese finger language makes the communication of the deaf and man-machine interaction more convenient. Traditional finger-language recognition adopts the method of convolution neural network(CNN) , leading to the structure of the model is single and a lot of information will be discarded in the pooling layer. Capsules are kinds of constructed and Abstract: d subnetworks in neural networks, and meanwhile each Capsule focuses on individual tasks and preserving spatial features of the image. Analyzing characteristics of finger language in Chinese sign language, and constructing and expanding training set of finger language pictures, we try to solve the task of finger language recognition by using CapsNet. Comparing the CapsNet recognition rate under different parameters and comparing with the classic GoogLeNet convolution network, experimental results show that CapsNet can achieve better recognition effect in the task of sign language recognition.
Hits
Hits
Downloads
Comment