6D pose perception facilitates the acquisition of precise position and orientation in three-dimensional space, thereby establishing a reliable foundation for intelligent sorting, robotic grasping control and automated packaging. However, existing 6D pose estimation methods encounter significant challenges in agricultural sorting lines. The reliance on instance-specific CAD models in conventional methods limits adaptability to natural variations in fruit scale and curvature, while the weak textures of smooth peel and dynamic inter-frame motion degrade feature extraction accuracy. Therefore, this study introduces a semi-dense image matching-driven 6D pose estimation framework, named SDM-Citrus6D, specifically designed for dynamic citrus pose tracking during conveyor inspection. The method first employs a segmentation model to achieve accurate detection and target mask extraction. A semi-dense image matching model, CitrusMatch, is then constructed, incorporating the Wavelet Feature Enhancement Module (WFEM) and the Adaptive Fuzzy Region Matching (AFRM) mechanism to enhance feature discriminability on weak-texture surfaces and suppress motion-blur interference. Based on the image matching results and depth information, a combination of the Kabsch and Iterative Closest Point (ICP) algorithms is applied to solve the 6D pose. A robotic experimental platform for fruit pose estimation was established to validate the proposed method. The results show that the CitrusMatch achieves a matching accuracy of Acc@5 px = 97.09
Accurate detection of greenhouse green peppers supports fruit counting, yield estimation, and crop monitoring. However, similarity between fruits and surrounding foliage, weak boundaries, occlusion, scale variation, and irregular fruit shapes hinder reliable detection. This study proposes a frequency-domain multi-scale dynamic and deformable detector (FMDD-YOLO) based on You Only Look Once version 11 (YOLOv11). Wavelet convolution is incorporated into the backbone to strengthen joint spatial- and frequency-domain representations of fruit structures and boundaries. A multi-scale cascaded attention module is placed after the spatial pyramid pooling stage to reduce fruit–background feature confusion and enhance responses to visible fruit boundaries. Content-aware dynamic upsampling improves cross-scale feature alignment, while a dynamic detection head with deformable convolution refines localization of curved, occluded, and scale-varying fruits. On the self-built greenhouse green pepper dataset, experiments with five random seeds yield a mean precision of 98.6
With the rapid advancement of financial technology and the increasing complexity and volatility of financial environments, traditional financial risk prediction methods face significant challenges in handling heterogeneous data, imbalanced risk labels, and limited scalability across different risk scenarios. These limitations hinder their effectiveness in identifying emerging risks in credit-card default, loan default, and credit-card fraud prediction. To address these challenges, this paper proposes FinGPT-Risk, a unified prompt-driven multi-task learning framework for financial risk prediction built upon the FinGPT large language model (LLM). Specifically, structured financial variables and textual profiles derived from those same variables are encoded and integrated within a shared representation space using task-aware prompts, enabling the model to capture transferable risk patterns and task-specific characteristics across multiple financial risk prediction sub-tasks. A task-specific attention mechanism is further introduced to adaptively emphasize critical risk signals, while a two-stage optimization strategy is adopted in which joint fine-tuning is followed by meta-training to improve task-level adaptation under heterogeneous data distributions. In addition, Bayesian optimization is used to determine task-specific decision thresholds, thereby improving the balance between precision and recall for imbalanced risk prediction. Extensive experiments on seven FinBench subsets covering credit-card default, loan default, and credit-card fraud demonstrate that FinGPT-Risk consistently outperforms classical machine learning, neural network, tabular learning, and LLM-based baselines. The proposed framework achieves average F1-scores of 39.95
Sparse Temporal Knowledge Graphs (STKGs) are characterized by severe data incompleteness, where missing observations not only reduce the available structural evidence but also progressively degrade entity representations and disrupt the temporal evolution of entities and relations. To address these coupled effects, this paper proposes RGSM-STKG, a progressive structural-temporal representation recovery framework for highly sparse temporal knowledge graphs. The framework first employs a dual-structure perception module to alleviate sparsity-induced entity representation degradation by jointly modeling local neighborhood information and relation-aware semantics, thereby providing more reliable historical entity-relation representations. These enhanced representations are then fed into a Mamba-based dynamic evolution module, which captures long-range temporal dependencies and reconstructs latent entity-relation states from incomplete historical observations. Finally, a wavelet-based multi-scale temporal inference module decomposes the reconstructed entity-relation state sequence along the temporal dimension to capture variations at different temporal scales. Multi-head attention then adaptively aggregates query-relevant historical context for target entity prediction. Experiments on eight sparse subsets constructed from ICEWS14 (2
Universities, school districts, and online-learning platforms now move student records, learning analytics, and behavioural traces across institutional and national borders as a matter of routine, yet the tools used to assess the risks of these transfers still score each transfer on its own and ignore the relations that tie institutions, users, datasets, providers, countries, vulnerabilities, and regulations together. PV-RiskNet is an explainable heterogeneous graph learning framework built for this setting. It represents the domain as a typed graph with seven node types and seven relation types, and couples that graph with privacy- and vulnerability-aware gating, risk-sensitive attention, multi-task prediction, calibrated uncertainty estimation, and meta-path-anchored explanations that come with mitigation recommendations. Evaluated against ten baselines on a curated semi-synthetic dataset, the framework reaches a macro-F1 of 90.9
