Pattern Recognition最新文献

{"title":"CEA-Net: A multi-modal model for corn disease classification with dynamic fusion and cross-layer connection mechanism","authors":"Haoyang Wang,&nbsp;Guoxiong Zhou,&nbsp;Guiyun Chen","doi":"10.1016/j.patcog.2025.112788","DOIUrl":"10.1016/j.patcog.2025.112788","url":null,"abstract":"<div><div>Corn is one of the most widely cultivated crops globally, yet it remains highly susceptible to a variety of diseases. With the rapid advancement of deep learning, image-based methods for corn disease classification have emerged and achieved promising results. However, many existing approaches still face challenges such as reliance on single-source information and limited feature extraction capacity. To address these issues, this paper proposes a multi-modal model named CEA-Net. First, we introduce a Cross-layer Connection Model (CCM) for image processing, which integrates multi-level wavelet blocks, VMamba, and Transformer components through a cross-layer connectivity mechanism. This design enhances spatial information reorganization and facilitates efficient feature extraction and reuse within the visual backbone network. Second, we propose an Efficient Dynamic Attention Fusion (EDAF) module for multi-modal feature fusion. EDAF dynamically modulates the contribution of each modality, emphasizing dominant sources while efficiently enhancing the representational capability of feature maps. Finally, we introduce Adaptive Adversarial Cross-Entropy Meta-learning (AACEM) for model pre-training. By combining meta-learning with sharpness-aware minimization and utilizing adaptive adversarial cross-entropy loss, AACEM improves both generalization and overall performance. Experimental results show that CEA-Net achieves an accuracy of 97.40%, outperforming networks such as EfficientViM and D2R by margins of 0.81%, 0.56%, 0.67%, and 0.55% across various metrics, demonstrating its significant practical value in corn disease management. Our code and dataset are available at: <span><span>https://github.com/yiyuynanodesu/CEA-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112788"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738047","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}

{"title":"Intra-modal consistency for image-text retrieval through soft-label distillation","authors":"Da Chen ,&nbsp;Yangtao Wang ,&nbsp;Yanzhao Xie ,&nbsp;Siyuan Chen ,&nbsp;Weilong Peng ,&nbsp;Maobin Tang ,&nbsp;Meie Fang ,&nbsp;C. L. Philip Chen ,&nbsp;Ping Li ,&nbsp;Wensheng Zhang","doi":"10.1016/j.patcog.2025.112817","DOIUrl":"10.1016/j.patcog.2025.112817","url":null,"abstract":"<div><div>Image-text retrieval (ITR) plays a crucial role in measuring the semantic association between different modalities. Most existing ITR efforts focus solely on cross-modal similarity but overlook the equally important intra-modal similarity (i.e., similarity between images and similarity between texts), leading to erroneous image-text matches. To address this issue, we propose <u><strong>I</strong></u>ntra-modal <u><strong>C</strong></u>onsistency for ITR through <u><strong>S</strong></u>oft-label <u><strong>D</strong></u>istillation (termed as ICSD), which cleverly leverages the intra-modal similarity (including encoder-derived intra-modal similarity and teacher-derived intra-modal similarity) to guide the cross-modal similarity. Firstly, during the online learning process, we (i) directly obtain the encoder-derived intra-modal similarity from image features generated from the image encoder and text features generated from the text encoder, and (ii) measure the cross-modal similarity based on those features from our well-designed Proximity Discrimination Module (PDM), which highly distinguishes the relationships between different features. During the offline calculation process, we pre-calculate the teacher-derived intra-modal similarity from those features generated by an image teacher model and text teacher model. Secondly, we devise Intra-modal Similarity Fusion (ISF) to organically combine both encoder-derived intra-modal similarity and teacher-derived intra-modal similarity to guide the learning of cross-modal similarity, developing the intra-modal guidance loss by means of knowledge distillation. Finally, we combine the traditional contrastive loss to jointly optimize the image-text matching process. Our approach is plug-and-play and can be easily applied to existing ITR models without changing their original architectures. Extensive experiments on ITR models and datasets verify our method can achieve much higher performance with lower model complexity than the state-of-the-art (SOTA) approaches. For example, we respectively outperform existing baselines (e.g., CLIP<span><math><msub><mrow></mrow><mrow><mi>V</mi><mi>i</mi><mi>T</mi><mo>−</mo><mi>B</mi><mo>/</mo><mn>32</mn></mrow></msub></math></span> and CLIP<span><math><msub><mrow></mrow><mrow><mi>V</mi><mi>i</mi><mi>T</mi><mo>−</mo><mi>L</mi><mo>/</mo><mn>14</mn></mrow></msub></math></span>) by over 18 % and 36 % in RSUM on MSCOCO. The code of this project is available at GitHub: <span><span>https://github.com/chd1516/ICSD</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112817"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145685567","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}

{"title":"Deep semi-supervised relation preserving learning model","authors":"Chenxi Tian,&nbsp;Lingling Li,&nbsp;Xu Liu,&nbsp;Licheng Jiao,&nbsp;Fang Liu,&nbsp;Shuyuan Yang","doi":"10.1016/j.patcog.2025.112793","DOIUrl":"10.1016/j.patcog.2025.112793","url":null,"abstract":"<div><div>Feature learning models aim to mine more useful features to complete downstream tasks more effectively. However, many algorithms have problems such as instability. In this paper, a deep semi-supervised relation preserving learning model (DSRPL) is proposed by embedding adaptive graph learning and category information to solve these problems. First, an autoencoder is used for feature learning to obtain nonlinear deep features. Then, the structural information of the original sample space is retained through adaptive graph learning to achieve the relationship between deep features and original data. Finally, a representation layer is embedded in the network, which takes advantage of semi-supervised learning while also utilizing category information to increase the compactness of samples within the same category. By this way, feature extraction and classification are integrated into one model for alternating learning can obtain more effective results. Compared with latest methods, DSRPL achieves more stable and more effective results.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112793"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738053","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}

{"title":"Cross-scale adaptive transformer with hierarchical feature synergy for aerial small object detection","authors":"Wenke Zhang,&nbsp;Mengmeng Liao","doi":"10.1016/j.patcog.2025.112822","DOIUrl":"10.1016/j.patcog.2025.112822","url":null,"abstract":"<div><div>Small object detection has long been a challenging task in computer vision due to limited pixel representation, extremely small scales, and complex scene variations. These challenges are particularly pronounced in high-resolution aerial imagery, where small objects refer to targets that appear as small-scale in images due to long shooting distances, characterized by limited pixel coverage and sparse feature representation. To address these issues, this paper proposes a novel object detection framework based on a Cross-Scale Adaptive Transformer and Hierarchical Feature Synergy. The framework introduces a Cross-Scale Adaptive Transformer module (CST) to dynamically capture multi-scale features in horizontal and vertical directions. Simultaneously, a Hierarchical Feature Synergy module (HFS) is designed to integrate low-, mid-, and high-level features, thereby enhancing semantic consistency and spatial detail preservation. Furthermore, we develop a novel loss function optimized for small object detection in aerial scenes, where small objects are caused by long shooting distances, effectively improving classification and localization accuracy. Extensive experiments on public datasets, including AI-TOD, VisDrone2019, and NWPU-VHR10, demonstrate that the proposed method significantly outperforms existing approaches in accuracy and efficiency. This work provides a new solution for practical aerial image analysis applications.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112822"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145685583","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}

{"title":"Structure and sensitivity in 3D human pose similarity quantification and estimation","authors":"Kyoungoh Lee ,&nbsp;Jungwoo Huh ,&nbsp;Jiwoo Kang ,&nbsp;Sanghoon Lee","doi":"10.1016/j.patcog.2025.112805","DOIUrl":"10.1016/j.patcog.2025.112805","url":null,"abstract":"<div><div>Recent advancements in deep learning have improved quantitative accuracy in 3D human pose estimation, but the estimated poses occasionally suffer from visual defects such as joint tremors and protrusions. While existing 3D pose similarity metrics and estimation models managed to reduce visual defects by addressing the structure of human poses, they still struggle in scenarios where visually sensitive joints are prevalent, particularly in cases of self-occlusion. In this paper, we identify these visually sensitive joints and demonstrate the significance of explicitly considering structure and sensitivity in the problem of 3D human pose estimation. Building upon the successful consideration of human pose structure, we first propose a new enhanced pose similarity metric PSIM<span><math><msup><mrow></mrow><mo>+</mo></msup></math></span>, which models sensitivity similarity to further capture human perception and focus on visual defects. Furthermore, we introduce a new 3D pose estimation model Dual Graph-based Convolutional Neural Networks (DG-CNN), which reconstructs 3D poses by focusing on the spatio-temporal correlation of the skeletal structure and actively controlling visually sensitive joints. By incorporating a novel similarity loss function, our model can implicitly model the structure and sensitivity of human poses through its architecture and explicitly through direct supervision. Our model not only improves the accuracy of the estimated pose but also increases the perceptual quality as evaluated by PSIM<span><math><msup><mrow></mrow><mo>+</mo></msup></math></span>, verifying the significance of structure and sensitivity awareness. Through rigorous benchmarking, we demonstrate that our metric and estimation model achieve the highest correlation with user scores and perform best in situations where visually sensitive joints are prevalent.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112805"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145685042","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}

{"title":"Multi-level ensemble feature selection for omics data","authors":"Xiaojian Ding ,&nbsp;Xin Wang ,&nbsp;Shilin Chen ,&nbsp;Kaixiang Wang","doi":"10.1016/j.patcog.2025.112809","DOIUrl":"10.1016/j.patcog.2025.112809","url":null,"abstract":"<div><div>High-dimensional, small-sample data poses substantial challenges for conventional feature selection, including instability, limited generalization, and vulnerability to noise. Ensemble feature selection offers a viable remedy, yet its effectiveness hinges on how base selectors are diversified and aggregated. To address this issue, we introduce a multi-level diversity framework that systematically enhances selector heterogeneity. At the sample level, density peak clustering is employed to identify structurally distinct low-dimensional embeddings, thereby strengthening separability. At the kernel level, diversity is quantified through the dispersion of kernel eigenvalues, which captures functional variation and improves representation robustness. By incorporating randomized feature mappings, we generate a pool of candidate projections that are further screened using the proposed metrics to form a diverse and reliable ensemble of feature selectors. We also designed three complementary aggregation strategies, namely EFS-RA (rank aggregation), EFS-SA (score aggregation), and EFS-IA (intersection aggregation), to derive a stable final feature subset. Comprehensive experiments on 15 real-world datasets show that our method consistently surpasses state-of-the-art approaches.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112809"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145685569","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}

{"title":"Query expansion with topic-aware in-context learning and vocabulary projection for open-domain dense retrieval","authors":"Ronghan Li ,&nbsp;Mingze Cui ,&nbsp;Benben Wang ,&nbsp;Yu Wang ,&nbsp;Qiguang Miao","doi":"10.1016/j.patcog.2025.112812","DOIUrl":"10.1016/j.patcog.2025.112812","url":null,"abstract":"<div><div>Large language models (LLMs) have recently emerged as pivotal components in open-domain dense retrieval. This study proposes a simple yet effective method to enhance dense retrieval using topic-aware In-Context Learning (ICL) and topic keyword projection. First we leverage LLMs to generate a pseudo-passage based on topic-aware demonstrations obtained from the pre-trained cluster to which the target query belongs. Second, we employ the masked language model (MLM) header in the autoencoder LM to map the query representation to implicitly topic-related tokens as keywords. We combine these two approaches to augment the original query. Extensive experiments on four prevalent open-domain question answering (ODQA) datasets demonstrate that our method achieves an average of 4.26 % improvement in R@20 compared to state-of-the-art query expansion work. Further analysis shows that the relevant demonstrations can provide higher-quality pseudo-passage generation, and the extracted keywords provide an interpretable basis for the effectiveness of dense retrieval. Code and data are available at <span><span>https://github.com/XD-BDIV-NLP/TDPR</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112812"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738128","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}

{"title":"Deep positional encoders for graph classification","authors":"Ahmed Begga,&nbsp;Miguel Ángel Lozano,&nbsp;Francisco Escolano","doi":"10.1016/j.patcog.2025.112828","DOIUrl":"10.1016/j.patcog.2025.112828","url":null,"abstract":"<div><div>Structural Pattern Recognition (SPR) includes the study of graphs as encoders of non-sequential and permutation-invariant patterns. In this regard, Graph Neural Networks (GNNs) are paving the way towards “inductive SPR” where classical structural problems such as graph classification can be approached through learnable priors. However, since graphs do not have a canonical order, existing GNNs struggle to learn the structural role of each node in the graph, which becomes key in graph classification. In this paper, we address this problem by making Spectral Graph Theory “inductive”, i.e. by learning the eigenvectors of the graph Laplacian, and then using them as positional encoders (PEs). Our experiments show that we improve significantly the SOTA of GNN-based graph classification.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112828"},"PeriodicalIF":7.6,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738054","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}

{"title":"Mambafusion: State-space model-driven object-scene fusion for multi-modal 3D object detection","authors":"Tong Ning,&nbsp;Ke Lu,&nbsp;Xirui Jiang,&nbsp;Jian Xue","doi":"10.1016/j.patcog.2025.112820","DOIUrl":"10.1016/j.patcog.2025.112820","url":null,"abstract":"<div><div>Existing multi-modal 3D detection struggles with geometric discrepancies between LiDAR/camera data and imbalanced feature alignment in Bird’s Eye View (BEV) space, where sparse foreground objects and scene-context gaps degrade performance. We propose MambaFusion, a novel framework unifying object-level fusion and scene-object interaction for robust 3D perception. Unlike scene-centric BEV fusion methods, MambaFusion introduces two modules: Object-Mamba, aligning 2D and 3D object candidates via grid-sorting and state-space models (SSM) to resolve modality inconsistencies, and Scene-Mamba, integrating image patches with object features and bidirectional SSM to model scene-object topological relationships. This dual-branch approach mitigates foreground-background imbalance and geometric misalignment while capturing holistic context. MambaFusion has achieved promising performance on both nuScenes and Waymo benchmarks.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112820"},"PeriodicalIF":7.6,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738056","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}

{"title":"FPQuant: A deep learning-based scalable framework for fingerprint phenomics quantification in large-scale biometric population studies","authors":"Zhiyong Han ,&nbsp;Yelin Shi ,&nbsp;Zhao Zhang ,&nbsp;Mu Li ,&nbsp;Haiguo Zhang ,&nbsp;Jingze Tan ,&nbsp;Wentian Zhen ,&nbsp;Tingting Liu ,&nbsp;Xueying Wang ,&nbsp;Chengyan Wang ,&nbsp;Jiucun Wang ,&nbsp;Li Jin ,&nbsp;Sijia Wang ,&nbsp;Manhua Liu ,&nbsp;Jinxi Li","doi":"10.1016/j.patcog.2025.112808","DOIUrl":"10.1016/j.patcog.2025.112808","url":null,"abstract":"<div><div>Fingerprint morphology, while evolutionary conserved yet individually distinct, emerges as a pivotal biometric identifier in anthropological research and forensic investigation. Current methodologies for precise identification and quantification of complex morphological features—particularly ridge counting and mean ridge-furrow pairs ridge breadth—remain constrained by labor-intensive and monolithic pattern recognition systems. This study presents FPQuant (Fingerprint Phenomics Quantification), a multi-task deep learning framework integrating the most comprehensive fingerprint pattern classification, singularity detection, and quantification of 12 morphometric phenotypes to date. Leveraging NSPT database of 28,867 expert-curated fingerprints, FPQuant achieved state-of-the-art performance with 97.18 % (6-class), 98.62 % (5-class), and 98.67 % (4-class) pattern classification accuracy; 98.63 % precision in topological singularity detection through optimized discrete keypoint localization; and expert-level precision in critical quantitative measurements including ridge counting. Cross-database validation demonstrated extraordinary generalizability with 96.20 % of 5-class accuracy on NIST-4 and 97.75 % of singularity precision on FVC2002 DB1. Notably, FPQuant’s integrated phenotypic capability revealed uncharacterized geographic variation in six morphometric traits, establishing novel fingerprint morphometric biomarkers for anthropological research. This study creates a scalable technical paradigm that bridging fingerprint phenomics with large-scale population study, while providing potential new research avenues across anthropology, forensics and biometric authentication.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"173 ","pages":"Article 112808"},"PeriodicalIF":7.6,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738125","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}