Tomato leaf diseases exhibit subtle inter-class differences and substantial intra-class variability, making accurate identification challenging for conventional deep learning models, especially under real-world conditions with diverse lighting, occlusion, and growth stages. Moreover, most existing approaches rely solely on visual features and lack the ability to incorporate semantic descriptions or expert knowledge, limiting their robustness and interpretability. To address these issues, we propose LPDiag, a multimodal prototype-attention diagnostic framework that integrates large language models (LLMs) for fine-grained recognition of tomato diseases. The framework first employs an LLM-driven semantic understanding module to encode symptom-aware textual embeddings from disease descriptions. These embeddings are then aligned with multi-scale visual features extracted by an enhanced Res2Net backbone, enabling cross-modal representation learning. A set of learnable prototype vectors, combined with a knowledge-enhanced attention mechanism, further strengthens the interaction between visual patterns and LLM prior knowledge, resulting in more discriminative and interpretable representations. Additionally, we develop an interactive diagnostic system that supports natural-language querying and image-based identification, facilitating practical deployment in heterogeneous agricultural environments. Extensive experiments on three widely used datasets demonstrate that LPDiag achieves a mean accuracy of 98.83%, outperforming state-of-the-art models while offering improved explanatory capability. The proposed framework offers a promising direction for integrating LLM-based semantic reasoning with visual perception to enhance intelligent and trustworthy plant disease diagnostics.

Addressing the challenges of variable target morphology, small critical regions, and complex background interference in eggplant picking point detection within complex agricultural scenarios, this study proposes MDAD-YOLO (Multi-dimensional Attention and DySample YOLO), a detection model improved based on the YOLOv10n-pose framework. First, the model’s cross-dimensional perception ability for fruits and picking points is enhanced by integrating the collaborative mechanism of regional receptive field attention with channel-space joint attention. Next, within the Neck structure, coordinate attention is incorporated to optimize the spatial localization accuracy of fine-grained features, enhancing sensitivity to minute regions such as the fruit stem apex. Additionally, dynamic pixel reorganization is applied to enhance feature map reconstruction details, addressing the detail loss caused by traditional interpolation methods. Finally, cascading adaptive fine-grained channel attention with position-sensitive attention enables multi-level modeling of channel dependencies and collaborative spatial context enhancement. Through a seven-tier validation framework, the model’s effectiveness, robustness, and generalizability have been comprehensively demonstrated. Experimental results show that the model achieves 93.6% mAP@50 for object detection, 94.7% mAP@50 and 92.1% mAP for keypoints detection, and an average pixel Euclidean distance error of 19.41 on the self-built eggplant dataset, outperforming YOLOv12 and other high-performance models. Additionally, cross-crop experiments on the pepper dataset showed a 2.1% and 2.7% improvement in mAP for object and picking point detection, respectively, compared to the baseline model, confirming its cross-crop robustness. This study reveals the synergistic enhancement of dynamic upsampling and attention mechanisms in agricultural object detection, providing new insights for lightweight model design in complex scenarios.

Learning reliable two-view correspondences is essential for geometric computer vision applications. Existing graph-based pruning methods typically aggregate information unidirectionally, capturing only whether a correspondence receives neighborhood support while ignoring its structural contribution during aggregation. This makes it difficult to distinguish locally clustered outliers from globally critical inliers. To overcome this limitation, we propose a Bidirectional Aggregation Graph Network (BAG-Net), which introduces a bidirectional aggregation perspective for correspondence pruning. Specifically, we design Bidirectional Dynamic Graph Convolution (BiDGC) that jointly models forward consistency aggregation and reverse structural dependency, enabling the network to identify key inliers that are widely referenced during feature propagation. To further enhance robustness, we design a Stage-wise Graph Expansion-Contraction (SGEC) architecture that combines multi-scale receptive field evolution with Stage-wise Squeeze-and-Excitation (SSE) and Deep-Guided Modulation (DGM) to stabilize feature propagation across network stages. Extensive experiments on the YFCC100M and SUN3D datasets demonstrate that BAG-Net significantly outperforms existing SOTA methods in both outlier rejection and camera pose estimation, achieving 66.10% and 78.88% mAP5^o% on YFCC100M in known and unknown scenes, respectively. Source code: https://github.com/LeyiWang13/BAG-Net.

Correspondence pruning aims to identify inliers from an initial set of correspondences with a low inlier ratio. Current Graph Neural Networks (GNNs) based correspondence pruning approaches suffer from feature over-smoothing during information propagation, making it difficult to distinguish inliers from outliers. In addition, Transformer-based methods can model long-range dependencies, but their quadratic complexity limits computational efficiency. To address these issues, we propose MatchMamba, a dual-view correspondence pruning network based on a selective state space model, Mamba. MatchMamba combines the strengths of GNNs and Mamba, enhancing local feature extraction while modeling global context with appropriate complexity. Specifically, to overcome Mamba’s limitations in correspondence pruning, such as the lack of local context and unidirectional modeling, we introduce the Cluster Sampling Spatial Mamba (CSSM) block and Correspondence Flip Bidirectional Mamba (CFBM) block. CSSM captures fine-grained local context through the implicit soft assignment and mitigates GNN’s over-smoothing using Mamba’s selective mechanism. CFBM block leverages Mamba’s efficient long-sequence modeling by constructing a pseudo-sequential structure through clustering. It applies forward and backward scanning to enable each correspondence to fully capture contextual information from others, achieving global context modeling with appropriate computational cost. Extensive experiments demonstrate that MatchMamba outperforms current state-of-the-art methods on several challenging tasks. The code is available at https://github.com/Mrwyb/MatchMamba.

Deep learning has shown remarkable potential in smart agriculture applications, particularly for crop disease diagnosis. However, the distributed nature of agricultural data, along with environmental heterogeneity and varying disease, poses critical challenges for data privacy, model reliability, and multi-terminal collaboration. To address these issues, this paper proposes a novel Federated Learning Uncertainty-Aware Aggregation framework (FLUA) with Paillier Homomorphic Encryption (PHE) for tomato leaf disease classification. FLUA integrates a global aggregation encoder with a local Uncertainty Estimation (UE) module based on evidential deep learning, which quantifies the reliability of each client. An Uncertainty-Aware Weighting (UAW) mechanism dynamically adjusts aggregation weights according to client-specific uncertainty scores, enabling robust model learning across heterogeneous data sources. Meanwhile, PHE ensures privacy-preserving communication and secure parameter aggregation among all participating clients. Experimental results on three heterogeneous tomato leaf disease datasets under realistic non-IID conditions demonstrate that FLUA outperforms state-of-the-art benchmarks, including FedProx and Moon, in terms of both classification accuracy and model reliability. Notably, the proposed approach improves trustworthiness in multi-institution collaborative training, effectively mitigating data privacy risks and enhancing robustness under distribution shifts. These findings highlight FLUA’s potential as a generalizable and secure federated learning paradigm for agricultural AI applications.

Robust models capable of maintaining strong generalization performance under data distribution shifts are crucial for real-world applications. Identifying flat minima is an effective strategy, especially the classic Sharpness-Aware Minimization (SAM). However, in domain generalization tasks, feature discrepancies of the same class across different domains can cause SAM’s uniform perturbation strategy to be dominated by domain-class pairs with large gradients, leading to inconsistent convergence and suboptimal sharp minima. To address this issue, we propose Class-Adaptive Variance-Guided Optimization (CAVGO). CAVGO adaptively calibrates perturbation directions by minimizing the variance of domain-class pair losses, achieving class-level cross-domain alignment. It further incorporates a novel gradient update strategy combined with Class-Adaptive Label Smoothing (CALS) to effectively mitigate bias toward hard-to-classify samples. By fully leveraging statistical differences across domains and classes, CAVGO significantly enhances model robustness and transferability under complex multi-source distributions. Extensive experiments on multiple domain generalization benchmarks demonstrate that CAVGO consistently outperforms existing methods, exhibiting stable generalization advantages and strong potential for practical applications.

Establishing reliable correspondences is a fundamental task in computer vision. Constructing neighbor graphs in feature space with position information to mine correspondence consistency has become a common strategy for recognizing correct correspondences (inliers). However, these neighbors may include a high ratio of incorrect correspondences (outliers), only using the correspondence consistency from feature space will probably be difficult to guarantee the matching accuracy. To address this issue, we propose a novel motion consistent space to find consistent neighbors that are independent of the correspondence's position and have a larger search range. On top of that, we build two neighbor graphs according to the feature space and motion consistent space separately, and expand a shift annular convolution to retain rich neighbor graph structure information and fully exploit the neighborhood context. Then, we design a dynamic feature fusion block to dynamically fuse these dual-neighbor graphs to flexibly cope with various complex scenarios. Finally, we develop a Dual-Neighborhoods Dynamic Fusion Network (DDFNet) for accurately identifying inliers and retrieving camera poses. Experimental results demonstrate that our proposed DDFNet outperforms the state-of-the-art methods. Source code: https://github.com/1211193023/DDFNet.

Whole slide image (WSI) classification is a critical task in computational pathology and is aimed at providing automated diagnostic support through high-resolution tissue image analysis. In weakly supervised WSI classification scenarios, the main challenge concerns the traditional multiple instance learning (MIL) methods, which rely on instance-level embeddings aggregated by an attention-based pooling mechanism. These methods often depend on data-driven statistical correlations, leading to misalignments between their attention allocation schemes and histopathological diagnostic regions and reducing the resulting prediction reliability. To address this, we propose frequency-aware causal regularized multiple instance learning (FC-MIL), an innovative framework combining that combines frequency-aware attention (FAA) and causal regularization (CR). FAA extracts more granular, fine-grained histological textures by jointly modeling spatial- and frequency- domain features, whereas CR introduces feature-level counterfactual perturbations as an intervention-inspired regularizer in the latent space, encouraging the model to rely less on spurious correlations and more on invariant pathological cues. Experimental results obtained on four WSI datasets show that FC-MIL outperforms the state-of-the-art MIL methods in terms of both accuracy and interpretability. Our source code is available at https://github.com/7FFDW/FCMIL.

Accurate identification of high-quality correspondences is essential for feature-based point cloud registration. However, relying solely on single-match associations to generate correspondences often yields a sparse set of inliers, which undermines the effectiveness of feature learning and fails to provide sufficient support for accurate transformation matrix estimation, particularly in low-overlap scenarios. Therefore, we propose a novel Gestalt-inspired bidirectional Parallel Interaction Network with inlier candidate expansion (GPI-Net++). It significantly increases the density and diversity of potential inliers, thereby promoting more effective feature interaction. Specifically, we introduce an Inlier Candidate Expansion (ICE) strategy, which leverages mutual top-k nearest neighbor matching to substantially improve inlier coverage in overlapping regions. To further refine the representation, an Orthogonal Integration (OI) strategy based on a projection mechanism is proposed to reduce redundancy and generate a compact and consistent structure from the candidate correspondences. Meanwhile, we incorporate a Dual-axis Attention Mining (DAM) block to extract geometric features from correspondences by exploring complementary information from correspondences and feature channels. Additionally, a Dual-path Multi-Granularity parallel interaction aggregation (DMG) block adopts a coarse-to-fine and fine-to-coarse strategy to promote cross-granularity feature interaction. Extensive experiments on various challenging tasks demonstrate better performance of our GPI-Net++ in comparison to existing methods. The code is released at https://github.com/gwk429/GPI-Net++.

Although traditional object detection has advanced, accurate recognition of small objects remains challenging for conventional detectors under computational efficiency constraints. Specifically, small objects usually generate weak and sparse feature responses, which cause detectors to allocate attention to dominant regions. In such cases, YOLO-based detectors often suffer from imbalanced attention allocation, resulting in unsatisfactory detection accuracy. To address these challenges, a novel small object detector called Progressive Modular-Guided Network (PMG-Net) is proposed in this paper. PMG-Net includes three innovative plug-and-play modules: Feature Aggregation and Enhancement (FAE) module, Large Kernel Multiscale Fusion (LKMF) module, and Dynamic Dual Pooling Attention Convolution (DDPAC) module. First, the FAE module strengthens fine-grained features through dual-branch serial multi-dimensional attention. Second, the LKMF module is embedded in shallow layers. It captures small object context by enhancing representation through large kernel decomposition and multiscale fusion. Finally, the DDPAC module enhances local and global collaboration by adaptively adjusting attention regions with learnable parameters and balancing weight allocation across feature regions. Experiments show that, compared with the YOLOv11-N baseline, PMG-Net improves AP by 2.9

Retrieving non-optically active water quality parameters from satellite imagery remains challenging. These variables lack distinct and directly interpretable spectral signatures. However, they usually require large labeled datasets.This study proposes a few-shot metric-learning framework, termed LFAB-DeepBDC, for continuous dissolved oxygen (DO) retrieval. The framework integrates a Local Feature Augmentation Block (LFAB) to enhance local feature representations. It also employs Deep Brownian Distance Covariance (DeepBDC) to capture second-order inter-channel dependencies. A prototype-weighted regression mechanism is further introduced to map metric similarities to continuous DO estimates.Experiments were conducted using Sentinel-2 imagery paired with in situ observations from 32 reservoirs across China. A strict temporal split was adopted. All training samples were acquired prior to January 1, 2025, while all testing samples were collected afterwards. LFAB-DeepBDC consistently outperformed DN4 and an empirical baseline. Relative errors were generally below 3%. In several cases, sub-percent relative errors were achieved.A long-term case study of Dahuofang Reservoir (2018–2025) further demonstrates stable retrieval behavior across seasonal regimes. The model also maintains reliable performance during episodic thermal and hydrologic disturbances. These results indicate that LFAB-DeepBDC provides a practical and data-efficient solution for satellite-based DO monitoring under dynamic environmental conditions.

Tea bud picking points detection is crucial for automating tea harvesting. However, existing detection algorithms still suffer from excessive parameter sizes and suboptimal accuracy. In addition, in natural environments, the variability in tea bud morphology and spatial distribution complicates detection. To address these challenges, this paper proposes a lightweight and high-precision deep learning method for tea bud picking point detection, based on the Real-Time DEtection Transformer (RT-DETR). The FasterNet block module is introduced to optimize the backbone of proposed network, significantly reducing computational complexity and parameters. Additionally, a novel Cross-Resolution Attention Fusion is employed, which adaptively combines fine-grained and coarse-grained features, enabling the integration of local and global information and improving robustness in complex environments. An improved loss function is also proposed, directing the model’s attention to hard samples and mitigating false negatives and positives. Experimental results show that our method outperforms the original RT-DETR, improving mean accuracy by 4.6

Road extraction from high-resolution remote sensing imagery remains an active research topic in the remote sensing and intelligent transportation communities. Despite recent advances, existing methods still face significant challenges, including fragmented road representations and inaccurate boundary delineation. This paper proposes a road extraction model, termed GLP-Mamba, built upon the RS-Mamba architecture. The proposed model introduces a Global–Local Perception (GLP) module and a Structural Consistency Enhancement (SCE) Block. GLP refines shallow feature representations before fusion, while SCE performs residual structural refinement during decoding. Together, they improve semantic discrimination, boundary representation and structural continuity. Experiments are conducted on three road extraction datasets: The DeepGlobe Roads Dataset, the Massachusetts Roads Dataset, and the Jilin-1 Satellite Dataset. The results show that GLP-Mamba achieves the highest F1-score and Intersection over Union (IoU) among the compared methods. Ablation studies further verify the effectiveness of each component. Specifically, the OSS Block provides long-range contextual modeling, the GLP module strengthens shallow feature representations before fusion, and the SCE Block helps refine road boundaries and reduce fragmented predictions. These results indicate that GLP-Mamba provides an effective solution for fine-grained road segmentation in complex remote sensing scenarios.

End-to-end imitation learning avoids hand-made robot motion for approaching and grasping, but the policy must still decide which fruit to pick and where to close the gripper. Occlusion can make the policy lose the selected fruit during harvesting, and the correct closing point is difficult to infer from pixels alone. This paper presents HarvestPoint-ACT, which makes both decisions explicit in perception and provides them to the policy. An instance segmentation front end with a keypoint branch predicts a mask and a harvest point for each visible fruit, where the harvest point specifies the location to close the gripper. A scheduler ranks detected candidates by occlusion and travel distance and selects one target. After each attempt, it redetects and reranks the candidates because the canopy may have changed. The selected fruit is encoded for an action chunking transformer as an eight-dimensional state, containing the absolute harvest point, the vector from the gripper to that point, a validity flag, and a confidence score. When the selected fruit is temporarily undetected, the system retains the last harvest point estimate in the robot base frame and marks it as stale, and aborts the attempt if the loss persists. On a canopy mock-up, HarvestPoint-ACT achieves a success rate of 88

The accuracy and stability of pathology image segmentation have become critical factors in clinical applications such as cancer screening and tumor grading. However, the presence of complex local structures, uncertain regions, and subtle morphological variations in pathological images continues to pose significant challenges. Most existing feature fusion approaches rely on the simplistic aggregation of extracted features, neglecting the unique characteristics and relative importance of distinct feature representations, which ultimately limits their potential to enhance model performance. To address these issues, we propose a Gestalt-Inspired Feature Integration Network (GeNet), a novel architecture inspired by Gestalt theory that mirrors the human visual system's ability to derive holistic understanding from partial information. Embracing the principle that ‘the whole is greater than the sum of its parts,’ GeNet introduces a mechanism to synergistically leverage multi-scale information, which assesses the similarity between features to achieve a more meaningful fusion of global context and local detail. Given the variability in target appearance within pathological images, we use information entropy to quantify feature uncertainty, allowing the model to prioritize uncertain regions and reduce the occurrence of ambiguous results. To explicitly eliminate multi-feature redundancy and misalignment, the refinement block utilizes parallel convolutional recalibration to fully leverage the advantages of various features. Extensive experiments on multiple pathological image segmentation datasets, including GlaS, GCaSeg, and EBHI-Seg, demonstrate that GeNet achieves high accuracy and strong robustness, offering a new perspective for joint modeling of global and local features in medical image analysis.

Low-resolution visible-infrared image fusion and super-resolution (LRVIF) are critical for enhancing image quality in low-resolution scenarios, yet limited information in the input images often constrains performance. To address these challenges, we propose SaDiff, a spatially-aware adaptive diffusion model that introduces diffusion processes into LRVIF for the first time, representing a major breakthrough in the field. Leveraging the generative capabilities of diffusion models, our approach unifies and enhances image fusion and super-resolution within a cohesive framework. A key component of SaDiff is the Spatial Residual Adaptation Block, which extends the diffusion process by dynamically adapting feature representations to spatial variations in the local regions of the input images. This module maximally preserves crucial information from the input images, such as texture details and contrast, while effectively suppressing noise, ensuring robust and context-aware feature refinement. Then we further propose Direct Diffusion Synthesis, a novel mechanism that utilizes noise predictions during diffusion to generate fused images, enabling joint training of the fusion and super-resolution networks. Additionally, a Cross-Feature Fusion Module integrates texture and contrast details, producing super-resolution fused images with improved clarity and structural integrity. Extensive experiments show that SaDiff achieves state-of-the-art performance, offering a robust and unified solution to infrared-visible image fusion and super-resolution. The code for the proposed method will be made available at https://github.com/guobaoxiao/SaDiff.

Correspondence pruning aims to identify inliers from an initial set of putative correspondences. Most existing methods establish putative correspondences via appearance-based matching and perform context aggregation directly on these pairs. While this effectively models cross-view dependencies, processing pairs as integral units overlooks the intra-view geometric structures formed by per-view keypoints. Consequently, such methods fail to leverage cross-view structural consistency, missing a critical geometric prior for separating inliers from outliers. To address this limitation, we propose the Hierarchical Geometric Consistency Network (HGCNet), which introduces a Hierarchical Geometric Relation (HGR) branch that directly leverages per-view keypoints to explicitly model intra-view geometric structures. Unlike existing methods that reason solely at the correspondence level, the HGR branch further exploits cross-view structural consistency as a complementary cue for verification, suppressing structurally inconsistent correspondence. Moreover, we introduce a Discrepancy-aware Enhancement (DaE) module and a Multi-Stage Attention Fusion (MSAF) module to preserve discriminative representations and effectively exploit hierarchical features across pruning stages. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance in two-view correspondence pruning. The code is available at https://github.com/ruifen98/HGCNet.

Automated needle-free injection can improve efficiency and biosecurity in precision livestock farming. However, robust robotic injection remains challenging because unpredictable pig motion disrupts target localization, injection accuracy, and safe physical interaction. To address these challenges, this paper presents a robotic needle-free injection system for moving pigs based on a vision-force hybrid servoing strategy. An RGB-D perception module enables real-time localization of the injection region and updates a desired injection point online. The injection point is then used to adjust the reference pose of an admittance controller, guiding the robot to track target motion while maintaining compliant and safe contact during injection. The system was quantitatively validated through 120 injections on a movable pig model and further validated in 60 injections on 20 live pigs. In live-animal experiments, it achieved a 93.3\% success rate with an average injection time of 6.8~s, while only two force-limit violations were observed. These results demonstrate that the proposed system can perform autonomous needle-free injection with high execution success and controlled interaction forces under the tested farm conditions, highlighting its potential for automated injection execution in commercial pig farming.