检索结果-内蒙古大学图书馆

Enhance the old representations' adaptability dynamically for exemplar-free continual learning

NEUROCOMPUTING 2025年 639卷

作者： Li, Kunchi Ding, Chaoyue Wan, Jun Yu, Shan Xiamen Univ Technol Sch Comp & Informat Engn Xiamen 361024 Fujian Peoples R China Xiamen Univ Technol Fujian Key Lab Pattern Recognit & Image Understand Xiamen 361024 Fujian Peoples R China Chinese Acad Sci Inst Automat Beijing 100190 Peoples R China

In continual learning, new data often falls outside the distribution of previous data. Since the old model is trained solely on past tasks and has not encountered the new data, its learned representations lack the adaptability needed to accommodate these new inputs. This mismatch creates a conflict between the learned representations and the concepts of new classes, resulting in a significant representation shift during model updates and exacerbating the issue of catastrophic forgetting. In this work, we presents a two-stage training framework to enhance the learned representations' adaptability dynamically for exemplar-free continual learning (EFCL), dubbed ERA-EFCL. The first stage involves synthesizing data for the old classes using DeepInversion and combining them with the new data to train an expanded module and a feature fusion network (FFN). The expanded module identifies critical features overlooked by the old model, while the FFN integrates the old representations with these newly discovered features. By incorporating additional information that distinguishes between the old and new classes, the FFN generates fused representations enriched with more transferable features for the old classes. This helps align the old representations with the new concepts, enhancing their adaptability and reducing the representation shift. Next, we introduce global task-wise knowledge distillation to ensure balanced knowledge transfer alongside other losses, thereby improving representation learning during the training of the new model. Furthermore, the new classifier is refined using a class-balanced training strategy. Extensive experiments demonstrate that ERA-EFCL achieves favorable results across four benchmark datasets. The code is available at: "https://***/CSTiger77/ ExemplarFreeCL".

关键词： Continual learning Enhance representation's adaptability Representation alignment Knowledge distillation

来源：评论

学校读者我要写书评

暂无评论

Joint radical embedding and detection for zero-shot Chinese character recognition

引用

pattern recognitION 2025年 161卷

作者： Luo, Guo-Feng Wang, Da-Han Zhang, Xu-Yao Lin, Zi-Hao Zhu, Shunzhi Xiamen Univ Technol Sch Comp & Informat Engn Fujian Key Lab Pattern Recognit & Image Understand Xiamen Peoples R China Chinese Acad Sci Inst Automat State Key Lab Multimodal Artificial Intelligence S Beijing Peoples R China Univ Chinese Acad Sci Sch Artificial Intelligence Beijing Peoples R China

Radical-based Zero-shot Chinese character recognition (ZSCCR) has attracted much attentions in recent years as radicals are common mid-level primitives that can bridge seen and unseen classes. However, while existing radical detection and sequence matching based methods have the difficulty of training a high-performance radical detection model due to the lacking of positional information of radicals in the character images, the attribute embedding based methods directly drop the radical detection procedure and overlooked the rich structural and spatial information of radicals in the visual perspective. In this paper, we proposed a novel joint radical embedding and detection method (JRED) for ZSCCR, which integrates the radical detection into the radical-level attribute embedding to explore and utilize the radicals' structural and spatial information for ZSCCR. Specifically, we maintain a learnable embedding vector for representing each radical and learn their representations without the positional information of radicals. The learned radical representations are regarded as the radical detectors to obtain the radical-level attributes by sliding the detector on the visual feature maps. The attributes are then used for the character representation by attribute embedding. Comprehensive experiments on various datasets show that the proposed JRED method can achieve the state-of-the-art performance, demonstrating the effectiveness of our proposed approach.

关键词： Zero-shot learning Chinese character recognition Joint radical embedding and detection

来源：评论

学校读者我要写书评

暂无评论

Attribute graph clustering via transformer and graph attention autoencoder

引用

INTELLIGENT DATA ANALYSIS 2025年第2期29卷 306-319页

作者： Weng, Wei Hou, Fengxia Gong, Shengchao Chen, Fen Lin, Dongsheng Xiamen Univ Technol Coll Comp & Informat Engn 600 Ligong Rd Xiamen 361024 Peoples R China Xiamen Univ Technol Fujian Key Lab Pattern Recognit & Image Understand Xiamen Peoples R China Xiamen Fuyun Informat Tech Co Xiamen Peoples R China

Graph clustering is a crucial technique for partitioning graph data. Recent research has concentrated on integrating topology and attribute information from attribute graphs to generate node embeddings, which are subsequently clustered using classical algorithms. However, these methods have some limitations, such as insufficient information inheritance in shallow networks or inadequate quality of reconstructed nodes, leading to suboptimal clustering performance. To tackle these challenges, we introduce two normalization techniques within the graph attention autoencoder framework, coupled with an MSE loss, to facilitate node embedding learning. Furthermore, we integrate Transformers into the self-optimization module to refine node embeddings and clustering outcomes. Our model can induce appropriate node embeddings for graph clustering in a shallow network. Our experimental results demonstrate that our proposed approach outperforms the state-of-the-art in graph clustering over multiple benchmark datasets. In particular, we achieved 76.3% accuracy on the Pubmed dataset, an improvement of at least 7% compared to other methods.

关键词： Attribute graph clustering transformer autoencoder node embedding

来源：评论

学校读者我要写书评

暂无评论

HC-GCN: hierarchical contrastive graph convolutional network for unsupervised domain adaptation on person re-identification

引用

MULTIMEDIA SYSTEMS 2023年第5期29卷 2779-2790页

作者： Chen, Si Xu, Bolun Zhang, Miaohui Yan, Yan Du, Xia Zhuang, Weiwei Wu, Yun Xiamen Univ Technol Sch Comp & Informat Engn Fujian Key Lab Pattern Recognit & Image Understand Xiamen 361024 Peoples R China Jiangxi Acad Sci Inst Energy Res Nanchang 330096 Peoples R China Xiamen Univ Sch Informat Fujian Key Lab Sensing & Comp Smart City Xiamen 361005 Peoples R China

The unsupervised domain adaptation (UDA) task on person re-identification (ReID) aims at spotting a person of interest under cross-camera by transferring the person knowledge learned from a labeled source domain to an unlabeled target domain. Recently, the contrastive loss provides an effective approach for UDA person ReID by comparing global features of the pedestrians. Generally, the fine-grained local features are favorable to distinguish the pedestrian appearance changes. However, the traditional contrastive loss-based UDA methods ignore the importance of local details and the relationship between the different granularities of features. To overcome this problem, we propose a hierarchical contrastive graph convolutional network, termed HC-GCN, for UDA person ReID. We first build an effective hierarchical graph model to learn the relationship between the global and local pedestrian features, where the local features are obtained by rough split and affine transformation. Moreover, we introduce the contrastive loss to suppress the pedestrian-irrelevant features, where the global and local contrastive losses are used. Experiments demonstrate that our method can achieve superior performance on the challenging Market-1501 and MSMT17 datasets.

关键词： Person re-identification Unsupervised domain adaptation Graph convolutional network Contrastive learning

来源：评论

学校读者我要写书评

暂无评论

Low-light image enhancement with quality-oriented pseudo labels via semi-supervised contrastive learning

引用

EXPERT SYSTEMS WITH APPLICATIONS 2025年 276卷

作者： Jiang, Nanfeng Cao, Yiwen Zhang, Xu-Yao Wang, Da-Han He, Yifan Wang, Chiming Zhu, Shunzhi Xiamen Univ Technol Sch Comp & Informat Engn Fujian Key Lab Pattern Recognit & Image Understand Xiamen Peoples R China Chinese Acad Sci Inst Automat State Key Lab Multimodal Artificial Intelligence S Beijing Peoples R China Univ Chinese Acad Sci Sch Artificial Intelligence Beijing Peoples R China Reconova Technol Co Ltd Xiamen Key Lab Visual Percept Technol & Applicat Xiamen Peoples R China

Existing Low-Light image Enhancement (LLIE) methods exhibit limitations when applied to complex scenes, primarily due to the lack of adaptation to real-world and synthetic images. To address this issue, we introduce a novel approach that leverages the mean-teacher strategy and contrastive regularization within a Contrastive Semi-Learning Network (CSL-Net). Specially, our CSL-Net employs high-quality pseudo labels to significantly improve model robustness and generalization. The novelties of this paper lie in: (1) We utilize a score- pool to store teacher's and student's outputs during training. After that, we propose a monotonic strategy to determine the most reliable Non-Reference image Quality Assessment (NR-IQA) metric, which provides effective evaluations and determines the optimal output as the pseudo-label. (2) We integrate this selected pseudo-label with a contrastive loss to provide auxiliary information that mitigates overfitting issues and guides student towards generating realistic images. (3) In network design, we incorporate illumination and gradient priors to refine color appearance and detail preservation, facilitated by several elaborated modules, including Contextual Refinement Blocks (CRBs), Non-Local Dynamic Attention (NLDA) and Illumination Perceptual Enhancer (IPE). Experimental results on popular datasets demonstrate that the enhancement quality can be significantly improved by our proposed method.

关键词： Low-light image enhancement Semi-supervised learning Contrastive learning

来源：评论

学校读者我要写书评

暂无评论

GridIIS: Grid Based Interactive image Segmentation 6th

GridIIS: Grid Based Interactive Image Segmentation

引用

6th Chinese Conference on pattern recognition and Computer Vision (PRCV)

作者： Zhu, Pengqi Wang, Da-Han Zhu, Shunzhi Xiaman Univ Technol Sch Comp & Informat Engn Xiamen Peoples R China Fujian Key Lab Pattern Recognit & Image Understan Xiamen Peoples R China

ISBN: (纸本)9789819985517;9789819985524

Interactive segmentation enables users to specify the object of interest (OOI) via various interaction strategies to obtain accurate segmentation results. An ideal interactive method should efficiently and accurately express users' segmentation intentions. However, the existing methods can only use a single interactive mode, ignoring the differences in scale and shape between OOIs, resulting in an inflexibility labeling process. In this paper, we propose a grid-based interactive image segmentation method (GridIIS). Specifically, GridIIS overlays grids on the image, and users can specify the location and shape of the OOI by selecting the grid areas as the interactive guidance. Users can choose the appropriate grid selection method and size considering the OOI's scale, shape, and boundary clarity to obtain guidance. We accordingly propose a novel grid sampling strategy, that considers the OOI's scale and shapes to adaptively estimate the grid size and area. Experiments on several datasets from different domains (street views, medical images, scene texts, etc.) show that our method achieves superior performance with fewer interaction rounds and exhibits strong generalization ability in cross-domain datasets.

关键词： Interactive segmentation Grid-based interactive

来源：评论

学校读者我要写书评

暂无评论

Bridge the Gap of Semantic Context: A Boundary-Guided Context Fusion UNet for Medical image Segmentation 7th

Bridge the Gap of Semantic Context: A Boundary-Guided Contex...

引用

7th Chinese Conference on pattern recognition and Computer Vision

作者： Chen, Yu Wu, Jiahua Wang, Da-Han Zhang, Xinxin Zhu, Shunzhi Xiamen Univ Technol Sch Comp & Informat Engn Xiamen 361024 Peoples R China Fujian Key Lab Pattern Recognit & Image Understan Xiamen 361024 Peoples R China

ISBN: (纸本)9789819784981;9789819784998

Accurate medical image segmentation of lesion areas is crucial in assisting diagnostics and treatment planning of diseases. In this paper, we propose a boundary-guided context fusion U-Net(BCF-UNet) for medical image segmentation to address a significant problem with existing approaches: ignoring the relationships between the boundary information and semantic context. In the BCF-UNet, we propose a novel feature fusion module, named Semantic Context-based Adaptive Fusion Module (SCAFM), to incorporate the different levels of encoder outputs and transfer them to the decoder. The SCAFM allows our network to fuse the deepest semantic information with the shallow edge information to greatly bridge the gap between different scale features. Then, in each layer of the encoder sub-network, we propose a newly pyramid edge enhancement module with edge-related prior knowledge to obtain the edge multi-granularity information, which enhances the accuracy of the segmentation boundary. In the BCF-UNet, edge information is peculiarly utilized to enhance the capability of semantic context extraction, which helps the model better understand and recognize the relationship between different tissue structures. The result of extensive experiments on two datasets shows that the proposed BCF-UNet outperforms stateof-art techniques.

关键词： Medical image Segmentation center dot Adaptive Feature Fusion center dot Edge Enhancement center dot U-Net

来源：评论

学校读者我要写书评

暂无评论

Stain-adaptive self-supervised learning for histopathology image analysis

引用

pattern recognitION 2025年 161卷

作者： Ye, Haili Yang, Yuan-yuan Zhu, Shunzhi Wang, Da-Han Zhang, Xu-Yao Yang, Xin Huang, Heguang Xiamen Univ Technol Sch Comp & Informat Engn Fujian Key Lab Pattern Recognit & Image Understand Xiamen 361024 Fujian Peoples R China Fujian Med Univ Union Hosp Dept Gen Surg 29 Xinquan Rd Fuzhou 350001 Fujian Peoples R China Chinese Acad Sci Inst Automat State Key Lab Multimodal Artificial Intelligence S Beijing Peoples R China Univ Chinese Acad Sci Sch Artificial Intelligence Beijing Peoples R China Huazhong Univ Sci & Technol Sch Elect Informat & Commun Wuhan 430000 Hubei Peoples R China

Staining variability is a critical factor affecting the accuracy of histopathological image analysis by reducing the distinguishability of tissue regions. Existing methods employ preprocessing techniques such as color matching and stain transfer for stain normalization, which can compromise data features. We propose a novel Stain-Adaptive Self-Supervised Learning (SASSL) method for histopathological image analysis. Our SASSL integrates a stain domain adversarial training module into the self-supervised learning (SSL) framework, allowing adaptation to staining variations while learning invariant features. SASSL can be viewed as a general invariant representation SSL method, with derived self-supervised weights applicable to various downstream tasks (classification, regression, and segmentation) in histopathological images. We conducted experiments on publicly available histopathological image analysis datasets, including PANDA, BreastPathQ, and CAMELYON16, achieving state-of-the-art performance. Results demonstrate that SASSL enhances feature extraction and mitigates the impact of staining variability, consistently improving performance across tasks. Our code is available at https://***/YeahHighly/SASSL_PR_2024.

关键词： Histopathology image analysis Domain adaptation Stain adaptation Self-supervised learning Domain adversarial training

来源：评论

学校读者我要写书评

暂无评论

UAM-Net: An Attention-Based Multi-level Feature Fusion UNet for Remote Sensing image Segmentation 6th

UAM-Net: An Attention-Based Multi-level Feature Fusion UNet ...

引用

6th Chinese Conference on pattern recognition and Computer Vision (PRCV)

作者： Cao, Yiwen Jiang, Nanfeng Wang, Da-Han Wu, Yun Zhu, Shunzhi Xiamen Univ Technol Sch Comp & Informat Engn Xiamen 361024 Peoples R China Fujian Key Lab Pattern Recognit & Image Understan Xiamen 361024 Peoples R China

ISBN: (纸本)9789819984619;9789819984626

Semantic segmentation of Remote Sensing images (RSIs) is an essential application for precision agriculture, environmental protection, and economic assessment. While UNet-based networks have made significant progress, they still face challenges in capturing long-range dependencies and preserving fine-grained details. To address these limitations and improve segmentation accuracy, we propose an effective method, namely UAM-Net (UNet with Attention-based Multi-level feature fusion), to enhance global contextual understanding and maintain fine-grained information. To be specific, UAM-Net incorporates three key modules. Firstly, the Global Context Guidance Module (GCGM) integrates semantic information from the Pyramid Pooling Module (PPM) into each decoder stage. Secondly, the Triple Attention Module (TAM) effectively addresses feature discrepancies between the encoder and decoder. Finally, the computation-effective Linear Attention Module (LAM) seamlessly fuses coarse-level feature maps with multiple decoder stages. With the corporations of these modules, UAM-Net significantly outperforms the most state-of-the-art methods on two popular benchmarks.

关键词： Semantic segmentation U-shape architecture Attention mechanism Feature fusion Remote sensing images

来源：评论

学校读者我要写书评

暂无评论

Graph Convolutional Networks Based on Neighborhood Expansion 7

Graph Convolutional Networks Based on Neighborhood Expansion

引用

7th International Conference on Electronic Information Technology and Computer Engineering (EITCE)

作者： Gong, Shengchao Weng, Wei Hou, Fengxia Lin, Dongsheng Wen, Juan Xiamen Univ Technol Xiamen Fujian Peoples R China Fujian Key Lab Pattern Recognit & Image Understan Xiamen Peoples R China Xiamen Univ Xiamen Fujian Peoples R China

ISBN: (纸本)9798400708305

Graph neural networks (GNNs) are an efficient framework for learning graph-structured data and achieving state-of-the-art performance on many tasks, including node classification, link prediction, and graph classi-fication. Over-smoothing problem is troublesome in GNNs, that is, increasing the depth of the network leads to significant degradation in model performance. A large number of methods address it through regu-larization or residual concatenation, but these methods only mitigate the over-smoothing in some extent, which persists as the network depth increases. In this paper, we propose the Graph Convolutional Network Based on Neighborhood Expansion (NEGCN), which enables the shallow network aggregate multi-hop neighbor features information for node embedding. specifically, NEGCN uses Depth-First-Search to ex-pand 1-hop neighbors with similarity, so as to avoid the problem of over-smoothing. The robustness of NEGCN is further improved by combining with attention mechanism and graph sparsification methods. Extensive experiments demonstrate that our model has a superior performance on real-world graph learning benchmarks.

关键词： Graph Convolutional Networks Graph Attention Networks Neighborhood expansion Over-smoothing

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：