检索结果-内蒙古大学图书馆

arXiv 2024年

作者： Xiong, Jiayu Wang, Jing Xiang, Hengjing Xue, Jun Xu, Chen Jiang, Zhouqiang Xiamen Key Lab of Computer Vision and Pattern Recognition Huaqiao University Xiamen China

Previous studies have highlighted significant advancements in multimodal fusion. Nevertheless, such methods often encounter challenges regarding the efficacy of feature extraction, data integrity, consistency of feature dimensions, and adaptability across various downstream tasks. This paper proposes a generalized multimodal fusion method (GMF) via the Poisson-Nernst-Planck (PNP) equation, which adeptly addresses the aforementioned issues. Theoretically, the optimization objective for traditional multimodal tasks is formulated and redefined by integrating information entropy and the flow of gradient backward step. Leveraging these theoretical insights, the PNP equation is applied to feature fusion, rethinking multimodal features through the framework of charged particles in physics and controlling their movement through dissociation, concentration, and reconstruction. Building on these theoretical foundations, GMF disassociated features which extracted by the unimodal feature extractor into modality-specific and modality-invariant subspaces, thereby reducing mutual information and subsequently lowering the entropy of downstream tasks. The identifiability of the feature’s origin enables our approach to function independently as a frontend, seamlessly integrated with a simple concatenation backend, or serve as a prerequisite for other modules. Experimental results on multiple downstream tasks show that the proposed GMF achieves performance close to the state-of-the-art (SOTA) accuracy while utilizing fewer parameters and computational resources. Furthermore, by integrating GMF with advanced fusion methods, we surpass the SOTA results. © 2024, CC BY-NC-SA.

关键词： Poisson equation

来源：评论

学校读者我要写书评

暂无评论

DocHFormer: Document Image Dewarping via Harmonized Modeling of Hierarchical Priors 27th

DocHFormer: Document Image Dewarping via Harmonized Modelin...

引用

27th International Conference on pattern recognition, ICPR 2024

作者： Zhou, Xinyue Li, Guanting Jiang, Nanfeng Wang, Da-Han Zhang, Xu-Yao Zhu, ShunZhi School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation of Chinese Academy of Sciences Beijing China

ISBN: (纸本)9783031781186

Document Image Dewarping (DID) task aims to address the issue of geometry distortion and improve image quality. In this paper, we propose a simple but effective method, named DocHFormer, that can take hierarchical priors features of images, including document image mask and coordinate positions, as additional information to realize accurate representation. To better exploit these fused information for dewarping, we take them into a harmonized space random shuffle operation, which can stochastically rearrange the pixels across spatial space and further use inverse operation to recover the original order. This way can adapt to allocate each feature pixel with equal probability and thus make full use of multi-type features. Furthermore, we introduce this mechanism into local self-attention to use linear complexity to input resolution and also design a new feed-forward network with structural modeling to boost representation. With the help of the above components, our proposed DocHFormer can achieve competitive performance with lower complexity and also outperform the existing state-of-the-art on several popular datasets. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Image enhancement

来源：评论

学校读者我要写书评

暂无评论

Document Image Shadow Removal via Frequency Information-Oriented Network 27th

Document Image Shadow Removal via Frequency Information-Ori...

引用

27th International Conference on pattern recognition, ICPR 2024

作者： Yang, Fan Zhou, Xinyue Jiang, Nanfeng Wang, Da-Han Zhang, Xu-Yao Li, Guantin Man, Wang Wu, Yun School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation of Chinese Academy of Sciences Beijing China

ISBN: (纸本)9783031781186

Removing shadows from document images can significantly improve the Quality of Experience (QoE) and boost the performance of the downstream document analysis and recognition tasks. However, existing methods still have limited generalization ability on complex document images and are prone to disrupt the image details. To address this issue, we consider the different shadow types that impact the image content on different frequency sub-bands. This motivates us to exploit frequency-domain information and further design a Frequency Information-oriented Deshadow Network (FID-Net). The proposed FID-Net mainly uses two elaborated modules, named Frequency Feature Extractor (FFE) and a Frequency Feature Refinement (FFR). FFE can generate low/high-frequency features through adaptively decomposing spectra of the shadow image. After that, FFR further refines both frequency features with mutual information operations. With the proposed key designs, extensive experimental results on the commonly used benchmarks demonstrate that the proposed method can learn discriminative shadows and achieve favorable performance against state-of-the-art approaches. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Image enhancement

来源：评论

学校读者我要写书评

暂无评论

A Dynamic Spatio-temporal Network with Self-attention for Multi-station Passenger Flow Prediction 12

A Dynamic Spatio-temporal Network with Self-attention for Mu...

引用

12th International Conference on Awareness Science and Technology, iCAST 2023

作者： Wang, Fengzhi Lv, Qinzhi Liu, Lijuan College of Computer and Information Engineering Xiamen University of Technology Xiamen China College of Computer and Information Engineering Xiamen University of Technology Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen China

ISBN: (纸本)9798350324693

Passenger flow prediction is vitally significant for intelligent transportation systems (ITS). Most of the studies typically focus on the passenger flow prediction for an individual station, and only capture the temporal features without considering any spatial features. Constructing a passenger flow prediction model for multiple stations, or even a whole network, is more valuable for practical applications. Therefore, we develop a dynamic spatio-temporal network (DSTNet) with a self-attention (SA) mechanism for multi-station passenger flow prediction. A dynamic graph convolutional network (DGCN) is applied for the spatial feature extraction, and gated recurrent unit (GRU) is combined to learn the temporal features. SA is applied to further assign the weights for the extracted spatio-temporal features. The Experiment has been conducted on the passenger flow in the xiamen bus rapid transit (BRT). The results demonstrate that the proposed DSTNet with SA (SA-DSTNet) outperforms the baselines in the multi-station passenger flow prediction task. © 2023 IEEE.

关键词： Forecasting

来源：评论

学校读者我要写书评

暂无评论

SANS: Spatial-Aware Neural Solver for Plane Geometry Problem 27th

SANS: Spatial-Aware Neural Solver for Plane Geometry Proble...

引用

27th International Conference on pattern recognition, ICPR 2024

作者： Lin, Zi-Hao Xiao, Shun-Xin Chen, Zi-Rong Li, Jian-Min Wang, Da-Han Zhang, Xu-Yao School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation of Chinese Academy of Sciences Beijing100190 China

ISBN: (纸本)9783031781186

Geometry problem solving (GPS) is an important research direction in artificial intelligence. Previous studies have demonstrated the effectiveness of neural solvers in GPS. However, they are deficiencies in accurately representing spatial relationships of geometric primitives within visually rich geometric diagrams. This paper presents a novel neural solver termed spatial-aware neural solver (SANS) that can perceive spatial relationships between geometric primitives. SANS includes two new modules: multimodal dual-branch spatial awareness pre-trained language module and point-primitive spatial-aware attention module. The pre-training module employs a dual-branch visual-textual point-matching strategy to align visual and textual points, and utilizes semantic structure pre-training to model global relationships. Additionally, the point-primitive spatial awareness attention module enhances the model’s ability to perceive spatial relationships between geometric primitives by accounting for the relative positions of points. Experiments show that SANS achieves 81.5 and 74.1 of accuracy on the Geometry3K and PGPS9K datasets. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Semantics

来源：评论

学校读者我要写书评

暂无评论

Learning Explicit Radical Representations for Zero-Shot Chinese Character recognition 27th

Learning Explicit Radical Representations for Zero-Shot Chin...

引用

27th International Conference on pattern recognition, ICPR 2024

作者： Pan, Song-Liang Wang, Da-Han Jiang, Nanfeng Zhang, Xu-Yao Zhu, Shunzhi School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation of Chinese Academy of Sciences Beijing100190 China

ISBN: (纸本)9783031781186

Zero-shot Chinese character recognition (ZSCCR) aims to recognize unseen Chinese characters by learning the semantic knowledge of seen characters. Radical-based methods treat Chinese characters as combinations of radicals, recognizing characters by predicting the radicals in the images. Existing radical-based methods have a closed radical parsing process that cannot be intervened in mid-course, relying only on semantic labels for constraints. However, semantic embedding vectors are usually manually designed and lack alignment with visual features, making it extremely difficult for the model to learn and locate discriminative radical representations from visual features. This paper proposes a ZSCCR network called Learning Explicit Radical Representations (LERRNet). LERRNet introduces learnable attribute hint vectors to guide the model in locating discriminative radicals and learning explicit representations of images. Specifically, we introduce a Radical Relevance Enhanced Encoder (RREE) to enhance the correlation of local radicals by augmenting the relationships between grid regions in visual features. Guided by attribute hint vectors, LERRNet employs a Radical Representation Decoder (RRD) to locate the most relevant regions of each radical in the given image and learn explicit radical representations. Extensive experiments demonstrate that LERRNet outperforms state-of-the-art radical/stroke-based methods across three ZSCCR benchmarks. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Zero-shot learning

来源：评论

学校读者我要写书评

暂无评论

WDA-Net: Weakly-Supervised Domain Adaptive Segmentation of Electron Microscopy

WDA-Net: Weakly-Supervised Domain Adaptive Segmentation of E...

引用

2022 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2022

作者： Qiu, Dafei Yi, Jiajin Peng, Jialin Huaqiao University Xiamen Key Laboratory of Computer Vision and Pattern Recognition China Meitu Inc Xiamen China Huaqiao University College of Computer Science and Technology China

ISBN: (纸本)9781665468190

Accurate segmentation of organelle instances is essential for electron microscopy analysis. Despite the outstanding performance of fully supervised methods, they highly rely on sufficient per-pixel annotated data and are sensitive to domain shift. Aiming to develop a highly annotation-efficient approach with competitive performance, we focus on weakly-supervised domain adaptation (WDA) with a type of extremely sparse and weak annotation demanding minimal annotation efforts, i.e., sparse point annotations on only a small subset of object instances. To reduce performance degradation arising from domain shift, we explore multi-level transferable knowledge through conducting three complementary tasks, i.e., counting, detection, and segmentation, constituting a task pyramid with different levels of domain invariance. The intuition behind this is that after investigating a related source domain, it is much easier to spot similar objects in the target domain than to delineate their fine boundaries. Specifically, we enforce counting estimation as a global constraint to the detection with sparse supervision, which further guides the segmentation. A cross-position cut-and-paste augmentation is introduced to further compensate for the annotation sparsity. Extensive validations show that our model with only 15% point annotations can achieve comparable performance as supervised models and shows robustness to annotation selection. © 2022 IEEE.

关键词： Electron microscopy

来源：评论

学校读者我要写书评

暂无评论

OCR4HSV: A Multi-task Learning Approach for Handwritten Signature Verification 27th

OCR4HSV: A Multi-task Learning Approach for Handwritten Sig...

引用

27th International Conference on pattern recognition, ICPR 2024

作者： Lin, Chao-Qun Wang, Da-Han Su, Yan-Fei Ge, De-Wu Zhang, Xu-Yao School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China Xiamen KEYTOP Communication Technology Co. Xiamen361024 China State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation of Chinese Academy of Sciences Beijing100190 China

ISBN: (纸本)9783031781186

Handwritten signature verification (HSV) models are notably recognized for their ability to discern whether a signature is forged in an offline document. Recently, HSV technology has made significant development. However, these methods have primarily focused on the shape features of the text image and overlooked the textual feature information inherent in the text itself, which makes the HSV model overfit. In this paper, we propose a novel network model named OCR4HSV to solve the above shortcomings. The proposed OCR4HSV first attempts to combine OCR and HSV to learn textual features in a multi-task learning manner. The model employs a dual-parameter approach, combining shared parameters and independent parameters. Specifically, Within the shared parameters, the Laplace attention module (LAM) is incorporated for edge information extraction. For independent parameters, CN-Mamba is utilized for sequence feature extraction in OCR, and the multi-scale global fusion block (MGFB) is designed to enhance the distinction between reference and test sample pairs. Leveraging OCR-related information and these architectural enhancements can fully mine the inherent textual feature information and significantly mitigate overfitting in the HSV task, thereby boosting verification accuracy. Our model has achieved state-of-the-art performance on the ChiSig and HanSig datasets. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

关键词： Multi-task learning

来源：评论

学校读者我要写书评

暂无评论

Tile selection method based on error minimization for photomosaic image creation

引用

Frontiers of computer Science 2021年第3期15卷 165-172页

作者： Hongbo ZHANG Xin GAO Jixiang DU Qing LEI Lijie YANG Department of Computer Science and Technology Huaqiao UniversityXiamen 361021China Fujian Key Laboratory of Big Data Intelligence and Security Huaqiao UniversityXiamen 361021China Xiamen Key Laboratory of Computer Vision and Pattern Recognition Huaqiao UniversityXiamen 361021China School of Computer Science and Technology Harbin Institute of TechnologyShenzhen 518055China

Photomosaic images are composite images composed of many small images called *** its overall visual effect,a photomosaic image is similar to the target image,and photomosaics are also called“montage art”.Noisy blocks and the loss of local information are the major obstacles in most methods or programs that create photomosaic *** solve these problems and generate a photomosaic image in this study,we propose a tile selection method based on error minimization.A photomosaic image can be generated by partitioning the target image in a rectangular pattern,selecting appropriate tile images,and then adding them with a weight *** on the principles of montage art,the quality of the generated photomosaic image can be evaluated by both global and local *** the proposed framework,via an error function analysis,the results show that selecting a tile image using a global minimum distance minimizes both the global error and the local error ***,the weight coefficient of the image superposition can be used to adjust the ratio of the global and local ***,to verify the proposed method,we built a new photomosaic creation dataset during this *** experimental results show that the proposed method achieves a low mean absolute error and that the generated photomosaic images have a more artistic effect than do the existing approaches.

关键词： photomosaic image tile image target image error minimization mean absolute error

来源：评论

学校读者我要写书评

暂无评论

Dlung:Unsupervised Few-Shot Diffeomorphic Respiratory Motion Modeling

引用

Journal of Shanghai Jiaotong university(Science) 2023年第4期28卷 536-545页

作者：陈培芝郭逸凡王大寒陈金铃 College of Computer and Information Engineering Xiamen University of TechnologyXiamen 361024FujianChina Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen 361024FujianChina School of Information Engineering Changchun Sci-Tech UniversityChangchun 130600China Department of Computer Science and Information Engineering Chaoyang University of TechnologyTaichung 41349TaiwanChina

Lung image registration plays an important role in lung analysis applications,such as respiratory motion *** learning-based image registration methods that can compute the deformation without the requirement of supervision attract much ***,it is noteworthy that they have two drawbacks:they do not handle the problem of limited data and do not guarantee diffeomorphic(topologypreserving)properties,especially when large deformation exists in lung *** this paper,we present an unsupervised few-shot learning-based diffeomorphic lung image registration,namely *** employ fine-tuning techniques to solve the problem of limited data and apply the scaling and squaring method to accomplish the diffeomorphic ***,atlas-based registration on spatio-temporal(4D)images is performed and thoroughly compared with baseline *** achieves the highest accuracy with diffeomorphic *** constructs accurate and fast respiratory motion models with limited *** research extends our knowledge of respiratory motion modeling.

关键词： unsupervised few-shot learning respiratory motion modeling diffeomorphic registration

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：