Accurate cattle pose estimation is important for precision livestock farming because body landmarks can provide quantitative visual inputs for potential health monitoring, behavior analysis, lameness assessment, and welfare evaluation. This study focuses on keypoint-localization performance and provides a foundation for future task-specific studies of these downstream outcomes. However, real farm images commonly contain inter-cattle occlusion, cluttered backgrounds, small anatomical landmarks, and visually similar animals, which limit the reliability of existing one-stage pose estimators. This study proposes LEC-Pose, a hierarchical locality refinement framework for occlusion-robust cattle pose estimation. Built on YOLOv8-Pose, LEC-Pose first predicts cattle boxes and auxiliary coarse keypoints, then extracts instance-level ROI features to recover local anatomical evidence. A lightweight refinement network directly predicts the final keypoints from heatmaps and offsets on enhanced ROI features. The detected boxes guide the ROI pathway, while the initial keypoints remain auxiliary outputs. During training, an instance-level contrastive loss regularizes a compact global ROI descriptor. During inference, pair construction and contrastive-loss computation are removed, while descriptor fusion remains in the prediction path. Across three runs on CattleEyeView, LEC-Pose reaches 39.47 ± 0.33 mAP@50:95, compared with 33.91 ± 0.28 for YOLOv8-Pose, while running at 111.6 FPS versus 138.2 FPS for YOLOv8-Pose, corresponding to a 19.2% throughput reduction on the evaluated RTX 4090. On NWAFU-Cattle, it obtains 76.08 ± 0.36 and 80.03 ± 0.38 mAP@50:95 under the 50%/50% and 80%/20% protocols, respectively; the former is comparable to FSMC-Pose, whereas the latter is the highest mean among the three repeated principal models. An end-to-end zero-shot evaluation on three external cattle datasets, second-dataset occlusion analysis, detailed module combinations, and descriptor diagnostics further assess the framework within their stated protocols. These results suggest that LEC-Pose can serve as a pose-estimation component for future automated cattle monitoring, and future task-specific studies can connect these keypoints to health, behavior, and welfare outcomes.

Accurate detection of blooming roses and flower buds is essential for greenhouse phenotyping, cultivation scheduling, harvest planning, and yield management. However, UAV-derived greenhouse imagery presents major challenges because rose targets are often small, densely distributed, partially occluded, and visually similar to complex backgrounds. This study proposes a lightweight rose detection framework that combines multi-source dataset construction with an improved YOLOv12n-based detector, termed BloomRoseNet. A GreenHouse Rose dataset was constructed by integrating self-collected UAV overhead images, screened RoseTracker images, and supplementary multi-view rose images to increase diversity in scale, growth stage, viewpoint, and background complexity. BloomRoseNet introduces task-oriented improvements for fine-grained feature extraction, adaptive feature fusion, and attention-enhanced detection. The supplementary multi-view data improved precision, recall, and mAP@50 from 85.2%, 82.8%, and 89.2% to 86.1%, 85.3%, and 90.5%, respectively. Compared with the baseline YOLOv12n, BloomRoseNet increased precision, recall, mAP@50, and mAP@50:95 by 3.2, 3.3, 3.6, and 1.6 percentage points, respectively, while reducing parameters from 2.55 M to 2.08 M and model size from 5.5 MB to 4.5 MB. The model also maintained real-time inference capability and stronger robustness under blur, occlusion, and illumination disturbances. The proposed framework provides an effective and practical solution for UAV-based greenhouse rose monitoring and supports precision cultivation management.

Aiming at the key technical challenges of complex background noise interference, fruit mutual occlusion, and multi-scale object recognition in natural scene Korla pear detection tasks, an improved YoloV11 object detection algorithm integrating the Efficient Channel Attention (ECA) mechanism and Bidirectional Feature Pyramid Network (BiFPN) is proposed (ECABiFPN-YOLOv11). By introducing the ECA module to adaptively optimize feature channel weights and combining the BiFPN architecture to achieve efficient cross-level feature fusion, the model’s perception and expression capabilities for multi-scale features of Korla pear objects are significantly enhanced. The experimental results show that the improved model reaches 86.8% on the mean average precision (mAP50) index, which is 4.7 percentage points higher than that of the original YoloV11 (82.1%). The mAP@0.5:0.95 value is 62.7%, which is 4.4% higher than that of the original model. The training box_loss (final) value is 3.7% lower than that of the original model, and the verification box_loss (final) value is 3.6% lower than that of the original model. These results provide reliable technical support for the research and development of automatic grading and sorting of fragrant pear fruits and intelligent picking systems in the field of smart agriculture.