Construction-scene object detection is complicated by background clutter, occlusion, large scale variation, and visually similar categories. We developed SFMDet, a YOLOv12-based detector that combines Selective Feature Scanning (SFS), Multi-level Cross-Attention (MCAM), Directional Partial Convolution (DPConv), and Context-Aware Decoupling (CADM). SFS models long-range spatial dependencies, whereas MCAM exchanges information across feature levels after spatial and channel alignment. DPConv limits the computational cost of neck fusion, and CADM supplies task-specific features for classification and localization. The Hybrid Adaptive Loss uses Varifocal Loss, WIoU-v1 with distribution focal loss, feature alignment, and epoch-dependent objective weights. On a construction-scene dataset containing 5,000 images, 49 categories, and 21,988 instances, SFMDet achieved 72.6 640×640 resolution and batch size 1. Evaluation on three public UAV and remote-sensing datasets showed gains over the same baselines. Code and dataset materials are available at https://github.com/XiaoxiaLuo/SFMDet .
Aiming at the difficulty of jointly addressing entity and event extraction, privacy risk, and communication efficiency in power-marketing text, this study proposes FedTrans-KIE, a federated Transformer framework for Edge IoT. The framework uses a shared Transformer encoder for joint entity recognition and event detection, applies client-side per-example gradient clipping with an L2 threshold of 1.0, performs bounded adaptive clipping of client updates, and adds Gaussian noise to the server-aggregated update before release. FedTrans-KIE achieved entity-recognition and event-detection F1 scores of 97.3
Abstract Testosterone Deficiency (TD), secondary TD in particular, is an endocrine disease which is very common in elderly men but often underdiagnosed during diagnosis. It consists of a range of physical, psychological, and metabolic disturbances that include loss of libido, fatigue, and depression, while the possibility of cardiovascular complications seems to increase among those with TD. Therefore, determination of secondary TD should be done genuinely and as early as possible so that the patient can be subjected to early remedial intervention and life improvement. However, conventional diagnostic methods mainly depend on symptomatic and isolated measurements of hormones, and these measures may, therefore, be insufficient for reliable detection. This study introduces an efficient and robust machine learning framework for the prediction of secondary testosterone deficiency using clinical and biochemical data. Preprocessing of data began with Min-Max normalization, then followed by random sampling to maintain scale consistency and class balance. A Relational Autoencoder (RAE) for the extraction of important features and to maintain the reduced dimensionality and subsequent optimized LightGBM classifier using histogram optimization (HO) technique for prediction with much better performance. The developed model represents a high-performance system with 96.61% accuracy, 95.86% precision, 96.19% recall, specificity of 96.92%, and an F1 score of 96.03%. Experimental evaluations exhibit that the proposed model significantly outperforms traditional existing models against the values of accuracy, precision, recall, specificity, and F1-score.
Abstract Long-running conversational agents must answer current questions from histories that can exceed a configured request budget. Sliding windows and rolling summaries reduce or rewrite history using rules that are not conditioned on the content of a later question. LiveCtx instead preserves the original sessions and selects context after the question arrives. Its Adaptive assembly policy uses the full history when it fits the budget; otherwise it retrieves complete selected sessions, leaves them unchanged when they fit, and compresses the selected set only once when it remains over budget. On 500 LongMemEval-S questions with Qwen3-Max generation and a common 131,072-character client budget, Adaptive achieved 79.8% and 81.8% accuracy when assessed independently by two automated graders. It exceeded Filter, which used the same retrieved sessions without generative compression, by 4.4 and 5.0 percentage points. The advantage was concentrated in the 342 over-budget retrieved sets and became small at the 160K-character budget, when 472 of 500 sets fit unchanged. In a fixed 50-question analysis, Adaptive used 42.2% fewer compression calls than per-session compression but retained fewer official target fragments. The evaluated policy therefore delayed lossy reduction until the question was known while keeping the original history available for later queries.
The Low Earth Orbit Satellite Internet of Things (LEO-SIoT), with its unique advantages, provides an ideal solution for achieving true ubiquitous connectivity. However, with the surge in terminal scale, spectrum resource scarcity and interference issues have become increasingly prominent. Spectrum cognition technology is an effective means to enhance spectrum utilization and avoid interference, but traditional ground-based equipment, due to fixed positions and limited coverage, often struggles to effectively perceive spectrum usage of dispersed satellite IoT terminals with skyward-pointing beams. To address this challenge, this paper proposes Sat-SCFramework, a UAV-assisted lightweight spectrum cognition framework leveraging UAV technology to achieve efficient spectrum cognition from an aerial perspective. We establish a refined scenario model describing UAV workflow for signal acquisition, image conversion, and data backhaul. Based on this, Sat-SCFramework accomplishes three typical spectrum cognition tasks: spectrum occupancy sensing, link quality assessment, and signal feature extraction. The core is Sat-SCformer, a novel lightweight deep learning algorithm for automatic classification of subcarrier modulation types. Sat-SCformer employs a dual-branch architecture, utilizing lightweight CNN to capture local features and efficient Transformer to mine global dependencies, achieving adaptive feature fusion and precise classification through multi-scale fusion and KAN modules. Finally, we build an experimental platform based on USRP X310, validating the framework through comprehensive and ablation experiments. Results demonstrate that Sat-SCformer achieves 98.25
