检索结果-内蒙古大学图书馆

arXiv 2025年

作者： Wu, Zhuguanyu Zhang, Jiayi Chen, Jiaxin Guo, Jinyang Huang, Di Wang, Yunhong State Key Laboratory of Virtual Reality Technology and Systems Beihang University China School of Computer Science and Engineering Beihang University Beijing China School of Artificial Intelligence Beihang University Beijing China

Vision Transformers (ViTs) have become one of the most commonly used backbones for vision tasks. Despite their remarkable performance, they often suffer significant accuracy drops when quantized for practical deployment, particularly by post-training quantization (PTQ) under ultra-low bits. Recently, reconstruction-based PTQ methods have shown promising performance in quantizing Convolutional Neural Networks (CNNs). However, they fail when applied to ViTs, primarily due to the inaccurate estimation of output importance and the substantial accuracy degradation in quantizing post-GELU activations. To address these issues, we propose APHQ-ViT, a novel PTQ approach based on importance estimation with Average Perturbation Hessian (APH). Specifically, we first thoroughly analyze the current approximation approaches with Hessian loss, and propose an improved average perturbation Hessian loss. To deal with the quantization of the post-GELU activations, we design an MLP Reconstruction (MR) method by replacing the GELU function in MLP with ReLU and reconstructing it by the APH loss on a small unlabeled calibration set. Extensive experiments demonstrate that APHQ-ViT using linear quantizers outperforms existing PTQ methods by substantial margins in 3-bit and 4-bit across different vision tasks. The source code is available at https://***/GoatWu/APHQ-ViT. Copyright © 2025, The Authors. All rights reserved.

关键词： Convolutional neural networks

来源：评论

学校读者我要写书评

暂无评论

A metadata extraction approach from papers based on meta-learning

引用

Journal of Information and Computational Science 2013年第4期10卷 1121-1129页

作者： Zhang, Fuzhi Zhao, Zihao School of Information Science and Engineering The Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Yanshan University Qinhuangdao 066004 China

To improve the accuracy of paper metadata extraction, a paper metadata extraction approach based on meta-learning is presented. Firstly, we propose a construction method of base-classifiers, which combines the Support Vector Machine (SVM) with the created diverse base-level training sets to construct some base-classifiers with larger diversity. The training sets are created according to the Open Access (OA) journal categories. Secondly, we present a paper metadata extraction algorithm based on meta-learning, which uses the meta-classifier to integrate the classification results of the base-classifiers and generates the final extraction results. The proposed approach guarantees extraction accuracy and independence of the base-classifiers, increases the diversities of them, and improves the classification performance of the meta-classifier, as well as reduces the degree of correlation of misclassifications. The experimental results show that our approach is superior to other single machine learning algorithm and the accuracy of paper metadata extraction is improved. © 2013 by Binary Information Press.

关键词： Support vector machines

来源：评论

学校读者我要写书评

暂无评论

VR content creation and exploration with deep learning: A survey

引用

Computational Visual Media 2020年第1期6卷 3-28页

作者： Miao Wang Xu-Quan Lyu Yi-Jun Li Fang-Lue Zhang State Key Laboratory of Virtual Reality Technology and Systems Beihang UniversityBeijing 100191China. Peng Cheng Laboratory Shenzhen 518000China School of Engineering and Computer Science Victoria University of WellingtonNew Zealand

virtual reality(VR)offers an artificial,computer generated simulation of a real life *** originated in the 1960 s and has evolved to provide increasing immersion,interactivity,imagination,and *** deep learning systems are able to represent and compose information at various levels in a deep hierarchical fashion,they can build very powerful models which leverage large quantities of visual media *** of VR methods and applications has been significantly boosted by the recent developments in deep learning *** content creation and exploration relates to image and video analysis,synthesis and editing,so deep learning methods such as fully convolutional networks and general adversarial networks are widely employed,designed specifically to handle panoramic images and video and virtual 3 D *** article surveys recent research that uses such deep learning methods for VR content creation and *** considers the problems involved,and discusses possible future directions in this active and emerging research area.

关键词： virtual reality deep learning neural networks 360°image and video virtual content

来源：评论

学校读者我要写书评

暂无评论

Discriminative feature encoding for intrinsic image decomposition

引用

Computational Visual Media 2023年第3期9卷 597-618页

作者： Zongji Wang Yunfei Liu Feng Lu Key Laboratory of Network Information System Technology(NIST) Aerospace Information Research InstituteChinese Academy of SciencesBeijing 100190China State Key Laboratory of Virtual Reality Technology and Systems School of Computer Science and EngineeringBeihang UniversityBeijing 100191China Peng Cheng Laboratory Shenzhen 518000China

Intrinsic image decomposition is an important and long-standing computer vision *** an input image,recovering the physical scene properties is *** physically motivated priors have been used to restrict the solution space of the optimization problem for intrinsic image *** work takes advantage of deep learning,and shows that it can solve this challenging computer vision problem with high *** focus lies in the feature encoding phase to extract discriminative features for different intrinsic layers from an input *** achieve this goal,we explore the distinctive characteristics of different intrinsic components in the high-dimensional feature embedding *** define feature distribution divergence to efficiently separate the feature vectors of different intrinsic *** feature distributions are also constrained to fit the real ones through a feature distribution *** addition,a data refinement approach is provided to remove data inconsistency from the Sintel dataset,making it more suitable for intrinsic image *** method is also extended to intrinsic video decomposition based on pixel-wise correspondences between adjacent *** results indicate that our proposed network structure can outperform the existing state-of-the-art.

关键词： intrinsic image decomposition deep learning feature distribution data refinement

来源：评论

学校读者我要写书评

暂无评论

Software fault localization via mining program dynamic execution graph

引用

Journal of Computational Information Systems 2014年第19期10卷 8563-8573页

作者： Zhang, Kai Yang, Ke Zhou, Lianbo Ren, Jiadong College of Information Science and Engineering Yanshan University Qinhuangdao China Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Qinhuangdao China

This paper is concerned with the problem of locating the code area related to software potential fault quickly and accurately in software testing period. A new method Sig BB based on graph model is proposed for mining the suspicious fault nodes from the passing and failing execution graphs. Representing each execution of a program as a graph, the graphs are divided into the passing and failing sets. By extracting the most representative passing and failing graphs based on these sets, the discriminative sub-graph is mined between the two representative graphs. First, Sig BB searches the max common graph, and then gets the opposite nodes set. The discriminative sub-graph is obtained by organizing and extending the set finally. Since the detected code scale is associated with the sorting of suspicious nodes, a suspicious metric strategy is also designed to sort the nodes in the discriminative sub-graph. Experimental results indicate that our method is both effective and efficient for software fault localization.

关键词： Software testing

来源：评论

学校读者我要写书评

暂无评论

A novel approach to efficiently mine structural patterns from software execution sequence

引用

Journal of Computational Information Systems 2015年第3期11卷 1109-1119页

作者： Huang, Guoyan Zhang, Bing Ren, Rong Ren, Jiadong College of Information Science and Engineering Yanshan University Qinhuangdao China The Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Qinhuangdao China

Analyzing structure of a software system is not sufficient for understanding its functionality. Certain characteristics of software are often hidden in its structure and can only be discovered when the various execution sequence of the software are identified. Analysis of execution sequence, therefore, becomes an indispensable aid in software comprehension. In this paper, a novel approach to efficiently mine structural patterns from software execution sequence is proposed. First, software execution traces are modeled as sequential patterns. Due to loops, patterns can occur multiple times within a trace, which leads to high cost of time and super complexity of the research. Then, an algorithm is designed to eliminate repetitive patterns (ERP) in software and software sequential pattern mining (SSPM) algorithm is presented to divide software unique sequential patterns into different patterns in order to mine structural characteristics of the software. At last, by the comparative analysis of similarity between different pattern collections, it draws the conclusion that software system can be compartmentalized into n fixed patterns. Experimental results also prove the conclusion well. ©, 2015, Binary Information Press. All right reserved.

关键词： computer software

来源：评论

学校读者我要写书评

暂无评论

DAU-Net: A medical image segmentation network combining the Hadamard product and dual scale attention gate

引用

Mathematical Biosciences and Engineering 2024年第2期21卷 2753-2767页

作者： Zhang, Xiaoyan He, Mengmeng Li, Hongan School of Computer Science and Technology Xi’an University of Science and Technology Xi’an710054 China State Key Laboratory of Virtual Reality Technology and Systems Beihang University Beijing100191 China

Medical image segmentation has an important application value in the modern medical field, it can help doctors accurately locate and analyze the tissue structure, lesion areas, and organ boundaries in the image, which provides key information support for clinical diagnosis and treatment, but there are still a large number of problems in the accuracy of the segmentation, so in this paper, we propose a medical image segmentation network combining the Hadamard product and dual-scale attention gate (DAU-Net). First, the Hadamard product is introduced in the structure of the fifth layer of the codec for element-by-element multiplication, which can generate feature representations with more representational capabilities. Second, in the jump connection module, we propose a dual scale attention gating (DSAG), which can highlight more valuable features and achieve more efficient jump connections. Finally, in the decoder feature structure, the final segmentation result is obtained by aggregating the feature information provided by each part, and decoding is achieved by up-sampling operation. Through experiments on two public datasets, Luna and Isic2017, DAU-Net is able to extract feature information more efficiently using different modules and has better segmentation results compared to classical segmentation models such as U-Net and U-Net++, and also verifies the effectiveness of the model. © 2024 American Institute of Mathematical Sciences. All rights reserved.

关键词： Image segmentation

来源：评论

学校读者我要写书评

暂无评论

New friend recommendation with user interest and socialization

引用

Journal of Information and Computational Science 2015年第11期12卷 4253-4262页

作者： Huang, Guoyan Chang, Yixin Li, Yakun College of Information Science and Engineering Yanshan University Qinhuangdao China The Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Qinhuangdao China

Friend recommendation is popular in social site to help people make new friends and expand their social networks. However, the conventional friend recommendation method is low accuracy for the sparsity of data and cold start. To solve this problem, we propose a new personalized recommendation method with user interest and socialization. This approach incorporating user interest and social information mainly leverages both user-based collaborative filtering method and user topological structure in social networks. Firstly, user interest degree can be obtained to identify the user preferences. Secondly, get the new optimized similarity degree and the target user's nearest neighbor by using social collaborative filtering algorithm. In the end, the final recommendation set of user can be acquired from the highest predicted rating of the candidate collection. The experiments indicate that this proposed method can enhance the Effectiveness and accuracy of recommendation compared with the traditional friend recommendation algorithm. ©, 2015, Binary Information Press. All right reserved.

关键词： Collaborative filtering

来源：评论

学校读者我要写书评

暂无评论

Road network information based edge bundling

引用

Journal of Information and Computational Science 2015年第13期12卷 4909-4918页

作者： Nie, Junlan Zhang, Jikai Xin, Shuyue Guo, Dongliang College of Information Science and Engineering Yanshan University Qinhuangdao China The Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Qinhuangdao China

To solve edge confusion problems in the geo-related graph visualization, here we present an edge bundling algorithm based on the road network information. First, the improved Dijkstra shortest path algorithm is employed to calculate the related path of original edges, and slicing quadratic Bezier curves are fitted out to replace original edges according to these path information. And then, the curve segments that satisfy conditions are merged together according to the control point consolidation strategy and thickened according to their weights, which simplifies the bundling result and also effectively shows the heat of each path;Finally, for the interested curves in the human-computer interaction, highlighted transition colors are introduced to character their directivity, and the effect is intuitive. The experimental results show that, our algorithm can effectively bundle original edges, which reduces the confusion degree of the graph. Meanwhile, the bundling result is more close to the original path, which improves the readability and understandability of the graph effectively. ©, 2015, Journal of Information and Computational Science. All right reserved.

关键词： Human computer interaction

来源：评论

学校读者我要写书评

暂无评论

Non-human-like characters' facial animation driven by one or two actors

引用

Journal of Information and Computational Science 2015年第10期12卷 3937-3944页

作者： Tang, Yong Liu, Yuhan Guo, Dan Liang, Haiyan Luo, Yanqing Lv, Mengya College of Information Science and Engineering Yanshan University Qinhuangdao China Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Qinhuangdao China

We present a novel system that can transfer facial expressions from an actor to a non-human-like character. Compared with a lot of previous studies, our system requires no special light conditions or facial markers for actors' performance. More significantly, our work can find vertex correspondences between the human face and characters with dissimilar facial structures by creating local-coordinate systems on the character. Experiments show that in our system not only one actor's performance is accurately transferred to non-human-like characters but also two actors' expressions can be simultaneously retargeted to one non-human-like character. Through our approach, virtual characters can have much more vivid expressions. 1548-7741/Copyright © 2015 Binary Information Press

关键词： Animation

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：