Pattern Recognition最新文献_第6页

Prompt-oriented and frequency-regularized schrödinger bridge for unpaired rain streaks and raindrops removal 即时导向和频率正则化schrödinger桥梁，用于不配对的雨条和雨滴去除

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-05 DOI: 10.1016/j.patcog.2025.112862

Yuanbo Wen , Jing Qin , Ting Chen , Tao Gao

The removal of rain streaks and raindrops is essential for improving image visibility. However, most existing methods rely on paired rainy and clean images, which are difficult to acquire in real-world scenarios. To this end, we propose prior-oriented and frequency-regularized Schrödinger bridge (PFSB) for rain streaks and raindrops removal with unpaired training. Specifically, we initially formulate unpaired image deraining as a Schrödinger bridge problem. Furthermore, we demonstrate the locally quasi-convexity of structural similarity, and employ the multi-scale structural similarity constraint (MSSC) to minimize the duality gap between the primal and dual problems, ensuring linear convergence of gradient flow while preserving textural details. Meanwhile, we develop a context-preserving consistency modulator (CCM) guide the derained output toward clean content, thereby retaining rain-irrelevant features. Moreover, we propose a domain-representative prompt protocol (DPP), which enforces the generated sample to eliminate rain-relevant information and maintain alignment with the clean domain. Additionally, we utilize Bayesian frequency-domain regularization (BFR) to balance spectral consistency with clean references and repulsion from rainy patterns. Extensive experiments demonstrate that our method surpasses the existing well-performing unpaired learning approaches in both fidelity and photo-realism.

去除雨纹和雨滴对于提高图像的可见度至关重要。然而，大多数现有的方法依赖于配对的雨和干净的图像，这在现实场景中很难获得。为此，我们提出了面向先验和频率正则化的Schrödinger桥（PFSB），用于无配对训练的雨条和雨滴去除。具体来说，我们最初将非配对图像去训练表述为Schrödinger桥问题。进一步，我们证明了结构相似性的局部拟凸性，并利用多尺度结构相似性约束（MSSC）最小化原问题和对偶问题之间的对偶差距，保证梯度流的线性收敛，同时保留纹理细节。同时，我们开发了一个上下文保持一致性调制器（CCM），将保留的输出引导到干净的内容，从而保留与雨无关的特征。此外，我们提出了一个域代表提示协议（DPP），该协议强制生成的样本消除与雨相关的信息并保持与干净域的一致性。此外，我们利用贝叶斯频域正则化（BFR）来平衡光谱一致性与干净参考和雨水模式的排斥。大量的实验表明，我们的方法在保真度和照片真实感方面都超过了现有的表现良好的非配对学习方法。

{"title":"Prompt-oriented and frequency-regularized schrödinger bridge for unpaired rain streaks and raindrops removal","authors":"Yuanbo Wen , Jing Qin , Ting Chen , Tao Gao","doi":"10.1016/j.patcog.2025.112862","DOIUrl":"10.1016/j.patcog.2025.112862","url":null,"abstract":"<div><div>The removal of rain streaks and raindrops is essential for improving image visibility. However, most existing methods rely on paired rainy and clean images, which are difficult to acquire in real-world scenarios. To this end, we propose prior-oriented and frequency-regularized Schrödinger bridge (PFSB) for rain streaks and raindrops removal with unpaired training. Specifically, we initially formulate unpaired image deraining as a Schrödinger bridge problem. Furthermore, we demonstrate the locally quasi-convexity of structural similarity, and employ the multi-scale structural similarity constraint (MSSC) to minimize the duality gap between the primal and dual problems, ensuring linear convergence of gradient flow while preserving textural details. Meanwhile, we develop a context-preserving consistency modulator (CCM) guide the derained output toward clean content, thereby retaining rain-irrelevant features. Moreover, we propose a domain-representative prompt protocol (DPP), which enforces the generated sample to eliminate rain-relevant information and maintain alignment with the clean domain. Additionally, we utilize Bayesian frequency-domain regularization (BFR) to balance spectral consistency with clean references and repulsion from rainy patterns. Extensive experiments demonstrate that our method surpasses the existing well-performing unpaired learning approaches in both fidelity and photo-realism.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112862"},"PeriodicalIF":7.6,"publicationDate":"2025-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145790683","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

R-FGDepth: Towards foundation models for recurrent depth learning with frequency-Guided initialization and refinement R-FGDepth：基于频率引导初始化和细化的循环深度学习基础模型

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-05 DOI: 10.1016/j.patcog.2025.112843

Zhaoxin Fan , Gen Li , Zhongkai Zhou

Self-supervised monocular depth estimation has seen remarkable progress with the advent of coarse-to-fine architectures and recurrent refinement frameworks. Despite their success, coarse-to-fine pipelines often depend on deep encoders and hierarchical upsampling, which introduce high computational overhead and propagate spatial inconsistencies. On the other hand, recurrent refinement models, such as R-MSFM and RAFM, suffer from suboptimal feature representations and limited capacity to capture fine-grained structures. In this work, we introduce R-FGDepth, a novel recurrent refinement framework enhanced with frequency-guided mechanisms to address these limitations. Our approach features three key innovations: (1) a Spatial-Semantic Hybrid Convolution Encoder that achieves an optimal trade-off between spatial detail preservation and semantic abstraction, (2) a Frequency-guided Global Depth Initialization Module that enforces global geometric consistency, and (3) a Frequency-guided Adaptive Depth Refinement Module that effectively enhances high-frequency structures, such as thin poles, traffic signs, and pedestrians. Extensive experiments on the KITTI and Cityscapes datasets demonstrate that R-FGDepth surpasses both coarse-to-fine and prior recurrent refinement methods, achieving state-of-the-art accuracy with competitive computational efficiency. Furthermore, qualitative evaluations underscore its ability to preserve object boundaries and generalize effectively across diverse domains. With its lightweight design and robust performance, R-FGDepth sets a new benchmark for real-world self-supervised depth estimation, advancing one step closer to the realization of foundation models in depth perception.

随着从粗到精的体系结构和循环细化框架的出现，自监督单目深度估计取得了显著的进展。尽管它们取得了成功，但粗到精的管道通常依赖于深度编码器和分层上采样，这会带来很高的计算开销并传播空间不一致性。另一方面，循环精化模型，如R-MSFM和RAFM，受到次优特征表示和捕获细粒度结构的有限能力的影响。在这项工作中，我们引入了R-FGDepth，这是一种新的循环细化框架，增强了频率引导机制，以解决这些限制。我们的方法有三个关键的创新：(1)一个空间-语义混合卷积编码器，实现了空间细节保存和语义抽象之间的最佳权衡；(2)一个频率引导的全局深度初始化模块，强制全局几何一致性；(3)一个频率引导的自适应深度细化模块，有效增强高频结构，如细杆、交通标志和行人。在KITTI和cityscape数据集上进行的大量实验表明，R-FGDepth超越了粗到精和先前的循环细化方法，以具有竞争力的计算效率实现了最先进的精度。此外，定性评估强调了其保留对象边界和有效地在不同领域进行推广的能力。R-FGDepth凭借其轻巧的设计和强大的性能，为现实世界的自监督深度估计设定了新的基准，向实现深度感知的基础模型又迈进了一步。

{"title":"R-FGDepth: Towards foundation models for recurrent depth learning with frequency-Guided initialization and refinement","authors":"Zhaoxin Fan , Gen Li , Zhongkai Zhou","doi":"10.1016/j.patcog.2025.112843","DOIUrl":"10.1016/j.patcog.2025.112843","url":null,"abstract":"<div><div>Self-supervised monocular depth estimation has seen remarkable progress with the advent of coarse-to-fine architectures and recurrent refinement frameworks. Despite their success, coarse-to-fine pipelines often depend on deep encoders and hierarchical upsampling, which introduce high computational overhead and propagate spatial inconsistencies. On the other hand, recurrent refinement models, such as R-MSFM and RAFM, suffer from suboptimal feature representations and limited capacity to capture fine-grained structures. In this work, we introduce <strong>R-FGDepth</strong>, a novel recurrent refinement framework enhanced with frequency-guided mechanisms to address these limitations. Our approach features three key innovations: (1) a Spatial-Semantic Hybrid Convolution Encoder that achieves an optimal trade-off between spatial detail preservation and semantic abstraction, (2) a Frequency-guided Global Depth Initialization Module that enforces global geometric consistency, and (3) a Frequency-guided Adaptive Depth Refinement Module that effectively enhances high-frequency structures, such as thin poles, traffic signs, and pedestrians. Extensive experiments on the KITTI and Cityscapes datasets demonstrate that R-FGDepth surpasses both coarse-to-fine and prior recurrent refinement methods, achieving state-of-the-art accuracy with competitive computational efficiency. Furthermore, qualitative evaluations underscore its ability to preserve object boundaries and generalize effectively across diverse domains. With its lightweight design and robust performance, <strong>R-FGDepth</strong> sets a new benchmark for real-world self-supervised depth estimation, advancing one step closer to the realization of foundation models in depth perception.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112843"},"PeriodicalIF":7.6,"publicationDate":"2025-12-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738052","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

iClickSeg: Interactive click segmentation for zero-shot cross-category 3D part segmentation iClickSeg：用于零镜头跨类别3D零件分割的交互式点击分割

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-04 DOI: 10.1016/j.patcog.2025.112816

Xing Yi , Liu Liu , Qiupu Chen , Li Zhang , Dan Guo

3D part segmentation is a crucial task for various applications, including robotics and shape analysis. Despite advancements in data-driven approaches, supervised methods heavily rely on annotated data, limiting their effectiveness in open-world scenarios and handling out-of-distribution test shapes. To address these challenges, we propose a novel interactive Click Segmentation (iClickSeg) method that achieves zero-shot cross-category 3D part segmentation via an iterative user interaction way. Specifically, our approach simulates user interactions through positive and negative clicks to guide the segmentation process, focusing on regions of interest and allowing for iterative refinement. To achieve this goal, we design a click sampling strategy learn shape-based prior information from point cloud data, enabling better feature encoding between points. Under the learned shape prior, the segmentation model can maintain the topology consistency and boost the performance with a simple PointNet++ network incorporation. For better refinement, we also present a post-processing strategy using outlier removal and heuristic click for obtaining the smooth segments. Extensive experiments on PartNet, PartNetE and S3DIS datasets demonstrate the superiority of iClickSeg over category-level segmentation methods and zero-shot methods. Inference tests on the AKB-48 data further validate the method’s effectiveness and practicality in real-world scenarios.

三维零件分割是包括机器人和形状分析在内的各种应用的关键任务。尽管数据驱动方法取得了进步，但监督方法严重依赖于注释数据，限制了它们在开放世界场景和处理分布外测试形状中的有效性。为了解决这些挑战，我们提出了一种新的交互式点击分割（iClickSeg）方法，该方法通过迭代的用户交互方式实现零镜头跨类别3D零件分割。具体来说，我们的方法通过积极和消极点击来模拟用户交互，以指导细分过程，专注于感兴趣的区域，并允许迭代改进。为了实现这一目标，我们设计了一种点击采样策略，从点云数据中学习基于形状的先验信息，从而实现点之间更好的特征编码。在学习到的形状先验条件下，该分割模型通过简单的PointNet++网络集成，可以保持拓扑一致性，提高分割性能。为了更好地细化，我们还提出了一种后处理策略，使用离群值去除和启发式点击来获得平滑段。在PartNet、PartNetE和S3DIS数据集上的大量实验表明，iClickSeg优于类别级分割方法和零射击方法。对AKB-48数据的推理测试进一步验证了该方法在实际场景中的有效性和实用性。

{"title":"iClickSeg: Interactive click segmentation for zero-shot cross-category 3D part segmentation","authors":"Xing Yi , Liu Liu , Qiupu Chen , Li Zhang , Dan Guo","doi":"10.1016/j.patcog.2025.112816","DOIUrl":"10.1016/j.patcog.2025.112816","url":null,"abstract":"<div><div>3D part segmentation is a crucial task for various applications, including robotics and shape analysis. Despite advancements in data-driven approaches, supervised methods heavily rely on annotated data, limiting their effectiveness in open-world scenarios and handling out-of-distribution test shapes. To address these challenges, we propose a novel <em>i</em>nteractive <em>Click Seg</em>mentation (<em>iClickSeg</em>) method that achieves zero-shot cross-category 3D part segmentation via an iterative user interaction way. Specifically, our approach simulates user interactions through positive and negative clicks to guide the segmentation process, focusing on regions of interest and allowing for iterative refinement. To achieve this goal, we design a click sampling strategy learn shape-based prior information from point cloud data, enabling better feature encoding between points. Under the learned shape prior, the segmentation model can maintain the topology consistency and boost the performance with a simple PointNet++ network incorporation. For better refinement, we also present a post-processing strategy using outlier removal and heuristic click for obtaining the smooth segments. Extensive experiments on PartNet, PartNetE and S3DIS datasets demonstrate the superiority of iClickSeg over category-level segmentation methods and zero-shot methods. Inference tests on the AKB-48 data further validate the method’s effectiveness and practicality in real-world scenarios.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112816"},"PeriodicalIF":7.6,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145790842","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

FDNet: High-frequency disentanglement network with information-theoretic guidance for multivariate time series forecasting 基于信息论指导的多变量时间序列预测高频解纠缠网络

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-04 DOI: 10.1016/j.patcog.2025.112810

Ao Hu , Liangjian Wen , Jiang Duan , Yong Dai , Dongkai Wang , Shudong Huang , Jun Wang , Zenglin Xu

Multivariate time series forecasting (MTSF) is crucial for decision-making in various domains but faces challenges due to the low signal-to-noise ratio (SNR) in real-world data. While frequency-domain methods have been employed to address this challenge, they often discard high-frequency components, assuming they are predominantly noise, thereby overlooking valuable short-term and event-driven information. To address this limitation, we propose a novel disentangled representation learning framework that separates high-frequency components into informative signals and noise using mutual information maximization and minimization strategies. We introduce the Frequency Distangle Network (FDNet), which integrates disentanglement with low- and high-frequency decomposition, gated neural networks, and variable relationship fusion to effectively preserve and utilize high-frequency signals. Extensive experiments on 12 real-world MTSF datasets demonstrate that FDNet significantly outperforms leading frequency-domain and time-domain baselines, highlighting the importance of leveraging rather than eliminating high-frequency information. The source code is publicly available at: https://github.com/aohu1105/FDNet.

多变量时间序列预测（MTSF）在各个领域的决策中起着至关重要的作用，但由于现实数据的低信噪比（SNR）而面临挑战。虽然已经采用频域方法来解决这一挑战，但它们通常会忽略高频组件，假设它们主要是噪声，从而忽略了有价值的短期和事件驱动信息。为了解决这一限制，我们提出了一种新的解纠缠表示学习框架，该框架使用互信息最大化和最小化策略将高频成分分离为信息信号和噪声。本文介绍了频率距离网络（Frequency Distangle Network, FDNet），该网络将解纠缠与低频和高频分解、门控神经网络和变量关系融合相结合，以有效地保留和利用高频信号。在12个真实MTSF数据集上进行的大量实验表明，FDNet显著优于领先的频域和时域基线，突出了利用而不是消除高频信息的重要性。源代码可以在：https://github.com/aohu1105/FDNet上公开获得。

{"title":"FDNet: High-frequency disentanglement network with information-theoretic guidance for multivariate time series forecasting","authors":"Ao Hu , Liangjian Wen , Jiang Duan , Yong Dai , Dongkai Wang , Shudong Huang , Jun Wang , Zenglin Xu","doi":"10.1016/j.patcog.2025.112810","DOIUrl":"10.1016/j.patcog.2025.112810","url":null,"abstract":"<div><div>Multivariate time series forecasting (MTSF) is crucial for decision-making in various domains but faces challenges due to the low signal-to-noise ratio (SNR) in real-world data. While frequency-domain methods have been employed to address this challenge, they often discard high-frequency components, assuming they are predominantly noise, thereby overlooking valuable short-term and event-driven information. To address this limitation, we propose a novel disentangled representation learning framework that separates high-frequency components into informative signals and noise using mutual information maximization and minimization strategies. We introduce the Frequency Distangle Network (FDNet), which integrates disentanglement with low- and high-frequency decomposition, gated neural networks, and variable relationship fusion to effectively preserve and utilize high-frequency signals. Extensive experiments on 12 real-world MTSF datasets demonstrate that FDNet significantly outperforms leading frequency-domain and time-domain baselines, highlighting the importance of leveraging rather than eliminating high-frequency information. The source code is publicly available at: <span><span>https://github.com/aohu1105/FDNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112810"},"PeriodicalIF":7.6,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738127","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

HyperPoint: Multimodal 3D foundation model in hyperbolic space HyperPoint：双曲空间中的多模态三维基础模型

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-04 DOI: 10.1016/j.patcog.2025.112800

Yiding Sun , Haozhe Cheng , Chaoyi Lu , Zhengqiao Li , Minghong Wu , Huimin Lu , Jihua Zhu

Self-supervised learning has made significant progress in Natural Language Processing and Computer Vision. Nevertheless, it encounters significant obstacles in the 3D domain, primarily due to the scarcity of available data and the considerable challenge of effectively capturing hierarchical structures. Current methods in Euclidean space suffer from feature distortion and fail to model the semantic hierarchies inherent in cross-modal data. These challenges motivate us to adopt hyperbolic space, which excels at capturing multi-scale relationships and preserving the geometric structure of complex data. In this paper, we propose HyperPoint, the first multi-modal 3D foundational model in hyperbolic space. By projecting cross-modal features such as 3D point cloud, 2D images, and text into hyperbolic space, we leverage its tree-like properties to encode semantic hierarchies with minimal distortion. Our method integrates generative and contrastive learning while leveraging multi-loss optimization to enhance feature diversity and consistency. HyperPoint achieves a new state-of-the-art in 3D representation learning, e.g., 96.1 % accuracy on ScanObjectNN, 94.1 % accuracy on 10w10s on ModelNet40. Our code is available at: https://github.com/Issac-Sun/HyperPoint.

自监督学习在自然语言处理和计算机视觉领域取得了重大进展。然而，它在3D领域遇到了重大障碍，主要是由于可用数据的稀缺性和有效捕获分层结构的相当大的挑战。现有的欧几里得空间方法存在特征失真的问题，无法对跨模态数据固有的语义层次进行建模。这些挑战促使我们采用双曲空间，它擅长捕捉多尺度关系和保持复杂数据的几何结构。本文提出了双曲空间中第一个多模态三维基础模型HyperPoint。通过将3D点云、2D图像和文本等跨模态特征投影到双曲空间中，我们利用其树状属性以最小的失真对语义层次进行编码。我们的方法结合了生成学习和对比学习，同时利用多损失优化来增强特征的多样性和一致性。HyperPoint实现了3D表示学习的新技术，例如，ScanObjectNN的准确率为96.1%，ModelNet40的10w10s的准确率为94.1%。我们的代码可在：https://github.com/Issac-Sun/HyperPoint。

{"title":"HyperPoint: Multimodal 3D foundation model in hyperbolic space","authors":"Yiding Sun , Haozhe Cheng , Chaoyi Lu , Zhengqiao Li , Minghong Wu , Huimin Lu , Jihua Zhu","doi":"10.1016/j.patcog.2025.112800","DOIUrl":"10.1016/j.patcog.2025.112800","url":null,"abstract":"<div><div>Self-supervised learning has made significant progress in Natural Language Processing and Computer Vision. Nevertheless, it encounters significant obstacles in the 3D domain, primarily due to the scarcity of available data and the considerable challenge of effectively capturing hierarchical structures. Current methods in Euclidean space suffer from feature distortion and fail to model the semantic hierarchies inherent in cross-modal data. These challenges motivate us to adopt hyperbolic space, which excels at capturing multi-scale relationships and preserving the geometric structure of complex data. In this paper, we propose HyperPoint, the first multi-modal 3D foundational model in hyperbolic space. By projecting cross-modal features such as 3D point cloud, 2D images, and text into hyperbolic space, we leverage its tree-like properties to encode semantic hierarchies with minimal distortion. Our method integrates generative and contrastive learning while leveraging multi-loss optimization to enhance feature diversity and consistency. <em>HyperPoint</em> achieves a new state-of-the-art in 3D representation learning, <em>e.g.</em>, <strong>96.1 %</strong> accuracy on ScanObjectNN, <strong>94.1 %</strong> accuracy on <em>10w10s</em> on ModelNet40. Our code is available at: <span><span>https://github.com/Issac-Sun/HyperPoint</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112800"},"PeriodicalIF":7.6,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738051","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

State transition difference prediction for deep reinforcement learning 深度强化学习的状态转移差分预测

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-04 DOI: 10.1016/j.patcog.2025.112824

Haotian Chi , Zhaogeng Liu , Xing Chen , Bohao Qu , Jifeng Hu , Yuan Jiang , Hechang Chen , Yi Chang

Deep reinforcement learning (DRL) has achieved remarkable success in sequential decision-making tasks such as video games, robotic control, and autonomous driving. State representation learning (SRL) offers a promising avenue to enhance reinforcement learning (RL) by extracting meaningful information from raw data, thereby boosting sample efficiency. However, most existing SRL methods focus on predicting future states, which limits their ability to fully leverage the information in the differences between consecutive states in RL sequences. These differences reflect the environment’s transition dynamics, which are crucial for effective decision-making. However, their categorical diversity makes them difficult to capture using a single mechanism. To overcome this limitation, we introduce a novel state representation learning approach for RL, state transition difference prediction (STDP). Specifically, we establish the STDP framework to forecast state differences, enabling a forward difference model to train two encoders: one that extracts state structures and another that captures the intrinsic relationship between state and action. Furthermore, we design two optional prediction targets within the STDP framework, thoroughly addressing the diversity of state transition differences to develop representations that embody the environment’s dynamics. Finally, we selectively integrate these representations into the value function and policy networks, providing the agent with comprehensive and relevant information for decision-making. Empirical results indicate that STDP improves sample efficiency in both online and offline settings compared to state-of-the-art methods. Additionally, we perform extensive analyses to validate the effectiveness and robustness of STDP.

深度强化学习（DRL）在视频游戏、机器人控制、自动驾驶等连续决策任务中取得了显著的成功。状态表示学习（SRL）通过从原始数据中提取有意义的信息，从而提高样本效率，为增强强化学习（RL）提供了一条有前途的途径。然而，大多数现有的SRL方法侧重于预测未来状态，这限制了它们充分利用RL序列中连续状态之间差异信息的能力。这些差异反映了环境的转变动态，这对有效决策至关重要。然而，它们的分类多样性使它们难以用单一机制捕获。为了克服这一限制，我们为强化学习引入了一种新的状态表示学习方法——状态转移差异预测（STDP）。具体来说，我们建立了STDP框架来预测状态差异，使前向差异模型能够训练两个编码器：一个提取状态结构，另一个捕获状态和动作之间的内在关系。此外，我们在STDP框架内设计了两个可选的预测目标，彻底解决了状态转换差异的多样性，以开发体现环境动态的表示。最后，我们有选择地将这些表征整合到价值函数和政策网络中，为智能体提供全面和相关的决策信息。实证结果表明，与最先进的方法相比，STDP提高了在线和离线设置的样本效率。此外，我们进行了广泛的分析来验证STDP的有效性和鲁棒性。

{"title":"State transition difference prediction for deep reinforcement learning","authors":"Haotian Chi , Zhaogeng Liu , Xing Chen , Bohao Qu , Jifeng Hu , Yuan Jiang , Hechang Chen , Yi Chang","doi":"10.1016/j.patcog.2025.112824","DOIUrl":"10.1016/j.patcog.2025.112824","url":null,"abstract":"<div><div>Deep reinforcement learning (DRL) has achieved remarkable success in sequential decision-making tasks such as video games, robotic control, and autonomous driving. State representation learning (SRL) offers a promising avenue to enhance reinforcement learning (RL) by extracting meaningful information from raw data, thereby boosting sample efficiency. However, most existing SRL methods focus on predicting future states, which limits their ability to fully leverage the information in the differences between consecutive states in RL sequences. These differences reflect the environment’s transition dynamics, which are crucial for effective decision-making. However, their categorical diversity makes them difficult to capture using a single mechanism. To overcome this limitation, we introduce a novel state representation learning approach for RL, state transition difference prediction (STDP). Specifically, we establish the STDP framework to forecast state differences, enabling a forward difference model to train two encoders: one that extracts state structures and another that captures the intrinsic relationship between state and action. Furthermore, we design two optional prediction targets within the STDP framework, thoroughly addressing the diversity of state transition differences to develop representations that embody the environment’s dynamics. Finally, we selectively integrate these representations into the value function and policy networks, providing the agent with comprehensive and relevant information for decision-making. Empirical results indicate that STDP improves sample efficiency in both online and offline settings compared to state-of-the-art methods. Additionally, we perform extensive analyses to validate the effectiveness and robustness of STDP.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112824"},"PeriodicalIF":7.6,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738067","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Fast online ℓ0 elastic net subspace clustering via a novel dictionary update strategy 基于字典更新策略的快速在线l0弹性网子空间聚类

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-03 DOI: 10.1016/j.patcog.2025.112829

Wentao Qu , Lingchen Kong , Linglong Kong , Bei Jiang

Driven by rapidly growing data volumes and increasing demands for real-time analysis, online subspace clustering has emerged as a valuable tool for processing dynamic data streams. However, existing online subspace clustering methods struggle to capture the complex and evolving distribution of such data due to rigid dictionary learning frameworks. In this paper, we propose a novel ℓ₀ elastic net subspace clustering model that integrates the ℓ₀ norm and the Frobenius norm to achieve the desirable block diagonal property. To enable dynamic adaptation, we further design a fast online alternating direction method of multipliers featuring an innovative dictionary update strategy based on support points–a compact set capturing the underlying data distribution. By selectively updating dictionary atoms guided by the support points, the proposed method dynamically adapts to shifting data characteristics, thereby enhancing adaptability and computational efficiency. Moreover, we provide rigorous convergence guarantees for the algorithm. Extensive numerical experiments demonstrate superior clustering accuracy and computational efficiency of our method, confirming its suitability for real-time and large-scale data processing tasks.

随着数据量的快速增长和实时分析需求的不断增长，在线子空间聚类已成为处理动态数据流的一种有价值的工具。然而，由于严格的字典学习框架，现有的在线子空间聚类方法难以捕获这些数据的复杂和不断变化的分布。在本文中，我们提出了一种新的l0弹性网子空间聚类模型，该模型集成了l0范数和Frobenius范数以获得理想的块对角性。为了实现动态适应，我们进一步设计了一种快速在线交替方向乘法器方法，该方法具有基于支撑点的创新字典更新策略-一个捕获底层数据分布的紧凑集。该方法通过在支撑点的引导下有选择地更新字典原子，动态适应数据特征的变化，提高了自适应性和计算效率。此外，我们还提供了严格的收敛性保证。大量的数值实验表明，该方法具有较高的聚类精度和计算效率，适用于实时和大规模的数据处理任务。

{"title":"Fast online ℓ0 elastic net subspace clustering via a novel dictionary update strategy","authors":"Wentao Qu , Lingchen Kong , Linglong Kong , Bei Jiang","doi":"10.1016/j.patcog.2025.112829","DOIUrl":"10.1016/j.patcog.2025.112829","url":null,"abstract":"<div><div>Driven by rapidly growing data volumes and increasing demands for real-time analysis, online subspace clustering has emerged as a valuable tool for processing dynamic data streams. However, existing online subspace clustering methods struggle to capture the complex and evolving distribution of such data due to rigid dictionary learning frameworks. In this paper, we propose a novel ℓ<sub>0</sub> elastic net subspace clustering model that integrates the ℓ<sub>0</sub> norm and the Frobenius norm to achieve the desirable block diagonal property. To enable dynamic adaptation, we further design a fast online alternating direction method of multipliers featuring an innovative dictionary update strategy based on support points–a compact set capturing the underlying data distribution. By selectively updating dictionary atoms guided by the support points, the proposed method dynamically adapts to shifting data characteristics, thereby enhancing adaptability and computational efficiency. Moreover, we provide rigorous convergence guarantees for the algorithm. Extensive numerical experiments demonstrate superior clustering accuracy and computational efficiency of our method, confirming its suitability for real-time and large-scale data processing tasks.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112829"},"PeriodicalIF":7.6,"publicationDate":"2025-12-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738050","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Curvi-Tracker: Curvilinear structure segmentation refinement by iterative tracking 曲线跟踪器：通过迭代跟踪进行曲线结构分割细化

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-03 DOI: 10.1016/j.patcog.2025.112797

Zhan Heng, Maurice Pagnucco, Erik Meijering, Yang Song

Curvilinear structures are ubiquitous in various domains, such as blood vessels in medical images or roads in satellite images. The automation of curvilinear structure segmentation is highly beneficial because of the laborious and error-prone process of manual annotation. Existing methods produce segmentation results with decent pixel-level performance, but still with presence of incorrect connectivity. To overcome the challenge, this paper proposes Curvi-Tracker, a novel refinement framework that improves initial coarse segmentation results by deploying tracker agents on detected foreground pixels. The proposed framework has two main components: a Direction-Net and a Forward-Net, which jointly guide the movement of trackers in order to track the curvilinear object. A Direction-Aware Multi-Label loss and a Stepwise Masked loss are proposed for accurate tracking of curvilinear structures. Experiments on public datasets of various curvilinear objects including retinal vessels, roads and pavement cracks demonstrate that the proposed method consistently improves the topological correctness of coarse segmentation results coarse segmentation results, averaging overall 10 % of improvement in all three topological metrics.

曲线结构在医学图像中的血管、卫星图像中的道路等各个领域都是无处不在的。由于手工标注的过程费力且容易出错，曲线结构分割的自动化是非常有益的。现有方法产生的分割结果具有不错的像素级性能，但仍然存在不正确的连接。为了克服这一挑战，本文提出了一种新的细化框架Curvi-Tracker，该框架通过在检测到的前景像素上部署跟踪代理来改善初始粗分割结果。该框架由两个主要部分组成：方向网和前向网，它们共同引导跟踪器的运动以跟踪曲线目标。为了精确跟踪曲线结构，提出了一种方向感知的多标签损失和一种逐步掩蔽损失。在包括视网膜血管、道路和路面裂缝在内的各种曲线对象的公共数据集上进行的实验表明，所提出的方法持续提高了粗分割结果的拓扑正确性，在所有三个拓扑指标上平均提高了10%。

{"title":"Curvi-Tracker: Curvilinear structure segmentation refinement by iterative tracking","authors":"Zhan Heng, Maurice Pagnucco, Erik Meijering, Yang Song","doi":"10.1016/j.patcog.2025.112797","DOIUrl":"10.1016/j.patcog.2025.112797","url":null,"abstract":"<div><div>Curvilinear structures are ubiquitous in various domains, such as blood vessels in medical images or roads in satellite images. The automation of curvilinear structure segmentation is highly beneficial because of the laborious and error-prone process of manual annotation. Existing methods produce segmentation results with decent pixel-level performance, but still with presence of incorrect connectivity. To overcome the challenge, this paper proposes Curvi-Tracker, a novel refinement framework that improves initial coarse segmentation results by deploying tracker agents on detected foreground pixels. The proposed framework has two main components: a Direction-Net and a Forward-Net, which jointly guide the movement of trackers in order to track the curvilinear object. A Direction-Aware Multi-Label loss and a Stepwise Masked loss are proposed for accurate tracking of curvilinear structures. Experiments on public datasets of various curvilinear objects including retinal vessels, roads and pavement cracks demonstrate that the proposed method consistently improves the topological correctness of coarse segmentation results coarse segmentation results, averaging overall 10 % of improvement in all three topological metrics.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112797"},"PeriodicalIF":7.6,"publicationDate":"2025-12-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738129","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Hierarchical order preserving spectral embedding 保持层次顺序的谱嵌入

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-03 DOI: 10.1016/j.patcog.2025.112758

Zhiguo Long , Yinghao He , Hua Meng , Tianrui Li

Data reconstruction, a key focus in data mining, aims to represent high-dimensional data in low-dimensional spaces while preserving structural integrity for downstream tasks. Spectral embedding methods are widely used for data reconstruction of diverse data structures. However, traditional density-based spectral embedding approaches face two limitations: (1) relying heavily on local structural information (e.g., distances between local neighbors) to characterize similarity, and (2) failing to distinguish different levels of distant relationships of data (known as hierarchical order relationships), potentially distorting original data structures. To address these issues, we propose Hierarchical Order Preserving Spectral Embedding (HORSE). HORSE combines local density estimation and similarities between subclusters to jointly capture local and global structures to improve the similarity measure. To better preserve hierarchical order relationships, HORSE introduces a quadruplet loss function based on hierarchical groups of subclusters to guide the reconstructed data to have a similar hierarchical order relationships with the original data. Experimental results on synthetic and real-world datasets demonstrate the effectiveness of our approach in both data reconstruction and clustering.

数据重构是数据挖掘中的一个重点，其目的是在低维空间中表示高维数据，同时保持下游任务的结构完整性。谱嵌入方法被广泛用于各种数据结构的数据重构。然而，传统的基于密度的谱嵌入方法面临两个局限性：(1)严重依赖局部结构信息（例如，局部邻居之间的距离）来表征相似性；(2)不能区分数据的不同级别的距离关系（称为层次顺序关系），可能会扭曲原始数据结构。为了解决这些问题，我们提出了层次保序谱嵌入（HORSE）方法。HORSE结合局部密度估计和子簇之间的相似性来共同捕获局部和全局结构，以改进相似性度量。为了更好地保持层次顺序关系，HORSE引入了基于子聚类层次分组的四重损失函数，引导重构数据与原始数据具有相似的层次顺序关系。在合成数据集和真实数据集上的实验结果证明了我们的方法在数据重构和聚类方面的有效性。

{"title":"Hierarchical order preserving spectral embedding","authors":"Zhiguo Long , Yinghao He , Hua Meng , Tianrui Li","doi":"10.1016/j.patcog.2025.112758","DOIUrl":"10.1016/j.patcog.2025.112758","url":null,"abstract":"<div><div>Data reconstruction, a key focus in data mining, aims to represent high-dimensional data in low-dimensional spaces while preserving structural integrity for downstream tasks. Spectral embedding methods are widely used for data reconstruction of diverse data structures. However, traditional density-based spectral embedding approaches face two limitations: (1) relying heavily on local structural information (e.g., distances between local neighbors) to characterize similarity, and (2) failing to distinguish different levels of distant relationships of data (known as <em>hierarchical order relationships</em>), potentially distorting original data structures. To address these issues, we propose Hierarchical Order Preserving Spectral Embedding (HORSE). HORSE combines local density estimation and similarities between subclusters to jointly capture local and global structures to improve the similarity measure. To better preserve hierarchical order relationships, HORSE introduces a quadruplet loss function based on hierarchical groups of subclusters to guide the reconstructed data to have a similar hierarchical order relationships with the original data. Experimental results on synthetic and real-world datasets demonstrate the effectiveness of our approach in both data reconstruction and clustering.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112758"},"PeriodicalIF":7.6,"publicationDate":"2025-12-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145790681","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Robust and flexible multi-view subspace clustering with nuclear norm 鲁棒灵活的核范数多视图子空间聚类

IF 7.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Pattern Recognition

Pub Date : 2025-12-03 DOI: 10.1016/j.patcog.2025.112804

Shaojun Shi , Yibing Liu , Canyu Zhang , Sisi Wang , Feiping Nie

Multi-view clustering technique utilizes the complementarity and consistency among different view features to divide the samples into different classes. Subspace learning garners considerable attention since it can explore the local structure in different dimensions. Although, multi-view subspace clustering algorithms have obtained remarkable performance, there are still some issues: 1) Nonlinear separable data sets cannot be exactly cut, which makes the flexibility be restricted; 2) The noise and outliers reduce the model robustness; 3) The clustering effectiveness is not outstanding. To solve these problems, this paper proposes a Robust and Flexible Multi-view Subspace Clustering with Nuclear Norm (RFMSC_NN), which integrates Multiple Kernel Learning (MKL) and Low-Rank Representation (LRR) within a cohesive framework. Specifically, firstly, projecting the linearly non-separable data to the Reproducing Kernel Hilbert Space (RKHS); Subsequently, learning a self-representation matrix to measure the similarity among samples; Then, by imposing the low rank constraint to reduce the noise interference; Next, adopting a self-weighted strategy to learn the weights of diverse views; Finally, using the k-means algorithm to obtain the clustering results. An alternate iteration optimization technique is employed to solve the model. Comprehensive experiments are conducted. The experimental results demonstrate enhanced clustering performance comparing with contemporary advanced multi-view clustering approaches.

多视图聚类技术利用不同视图特征之间的互补性和一致性，将样本划分为不同的类。子空间学习由于能够在不同的维度上探索局部结构而备受关注。虽然多视图子空间聚类算法已经取得了显著的性能，但仍然存在以下问题：1)非线性可分数据集不能精确切割，限制了算法的灵活性；2)噪声和异常值降低了模型的鲁棒性；3)聚类效果不突出。为了解决这些问题，本文提出了一种鲁棒灵活的核范数多视图子空间聚类方法（RFMSC_NN），该方法将多核学习（MKL）和低秩表示（LRR）融合在一个内聚框架内。具体而言，首先，将线性不可分数据投影到再现核希尔伯特空间（RKHS）；随后，学习一个自表示矩阵来度量样本之间的相似性；然后，通过施加低秩约束来降低噪声干扰；其次，采用自加权策略来学习不同观点的权重；最后，利用k-means算法得到聚类结果。采用交替迭代优化技术对模型进行求解。进行了综合实验。实验结果表明，与当前先进的多视图聚类方法相比，该方法的聚类性能有所提高。

{"title":"Robust and flexible multi-view subspace clustering with nuclear norm","authors":"Shaojun Shi , Yibing Liu , Canyu Zhang , Sisi Wang , Feiping Nie","doi":"10.1016/j.patcog.2025.112804","DOIUrl":"10.1016/j.patcog.2025.112804","url":null,"abstract":"<div><div>Multi-view clustering technique utilizes the complementarity and consistency among different view features to divide the samples into different classes. Subspace learning garners considerable attention since it can explore the local structure in different dimensions. Although, multi-view subspace clustering algorithms have obtained remarkable performance, there are still some issues: 1) Nonlinear separable data sets cannot be exactly cut, which makes the flexibility be restricted; 2) The noise and outliers reduce the model robustness; 3) The clustering effectiveness is not outstanding. To solve these problems, this paper proposes a Robust and Flexible Multi-view Subspace Clustering with Nuclear Norm (RFMSC_NN), which integrates Multiple Kernel Learning (MKL) and Low-Rank Representation (LRR) within a cohesive framework. Specifically, firstly, projecting the linearly non-separable data to the Reproducing Kernel Hilbert Space (RKHS); Subsequently, learning a self-representation matrix to measure the similarity among samples; Then, by imposing the low rank constraint to reduce the noise interference; Next, adopting a self-weighted strategy to learn the weights of diverse views; Finally, using the k-means algorithm to obtain the clustering results. An alternate iteration optimization technique is employed to solve the model. Comprehensive experiments are conducted. The experimental results demonstrate enhanced clustering performance comparing with contemporary advanced multi-view clustering approaches.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112804"},"PeriodicalIF":7.6,"publicationDate":"2025-12-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145790677","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0