Accurate EMA assessment in live sheep is crucial for genetic breeding and production management within the meat sheep industry. However, the segmentation accuracy and reliability of existing automated methods are limited by challenges inherent to B-mode ultrasound images, such as low contrast and noise interference. To address these challenges, we present MPG-SwinUMamba, a novel deep learning-based segmentation network. This model uniquely combines the state-space model with a U-Net architecture. It also integrates an edge-enhancement multi-scale attention module (MSEE) and a pyramid attention refinement module (PARM) to improve the detection of indistinct boundaries and better capture global context. The global context aggregation decoder (GCAD) is employed to precisely reconstruct the segmentation mask, enabling automated measurement of the EMA. Compared to 12 other leading segmentation models, MPG-SwinUMamba achieved superior performance, with an intersection-over-union of 91.62% and a Dice similarity coefficient of 95.54%. Additionally, automated measurements show excellent agreement with expert manual assessments (correlation coefficient r = 0.9637), with a mean absolute percentage error of only 4.05%. This method offers non-invasive and efficient and objective evaluation of carcass performance in live sheep, with the potential to reduce measurement costs and enhance breeding efficiency.

Deep learning-based sheep face recognition technology significantly enhances the automation of individual sheep identification, providing critical technical support for smart livestock farming and precision agriculture. However, in real farming environments, factors such as complex backgrounds, illumination variations, and the high visual similarity of sheep faces severely constrain the comprehensive performance of recognition systems regarding accuracy and real-time capability. To address these challenges, we propose a cascaded framework comprising the WRT-DETR model for detection and LG-MobileViT for identification. WRT-DETR integrates multi-scale wavelet residual modeling and adaptive feature interaction into the RT-DETR architecture to effectively handle complex backgrounds. Subsequently, LG-MobileViT utilizes local–global collaborative modeling to distinguish fine-grained features while maintaining a lightweight footprint suitable for edge devices. Experiments conducted on a dataset of 400 individuals and 20,000 images demonstrate that WRT-DETR achieves 92.5% mAP50 in detection tasks. Furthermore, LG-MobileViT attains 98.97% recognition accuracy with a parameter size of only 4.57 MB. On edge computing platforms, the integrated system reaches an inference speed approaching 100 FPS. These results confirm that the proposed framework offers an efficient, reliable technical solution for non-contact, precise sheep identification in practical precision agriculture scenarios.

Point cloud semantic segmentation is a pivotal technology for realizing non-contact body measurement and refined management of livestock. However, processing sheep point clouds in smart livestock scenarios presents specific challenges, primarily due to non-rigid posture deformations and severe background interference. To address these issues, this paper proposes a novel symmetric encoder–decoder architecture named Local–Global Aggregation Network (LGA-Net), which achieves high-precision parsing of sheep point clouds by constructing a dual-scale feature aggregation mechanism. First, a Dual Attention Aggregation (DAA) module is designed to jointly encode geometric and color features, significantly enhancing the network’s ability to capture fine-grained local boundaries, such as sheep ears and hooves. Second, a Global Semantic Relation (GSR) module is introduced, utilizing spatial occupancy ratios to establish long-range dependencies, thereby effectively resolving semantic ambiguity caused by posture variations. Furthermore, a plug-and-play Dual-domain Feature Enhancement (DFE) module is proposed. By fusing bilinear interactions between explicit 3D space and implicit feature space, the DFE module constructs a high-pass filtering mechanism to suppress low-frequency background noise. Extensive experiments on a self-constructed point cloud dataset involving two semantic classes (Sheep and Fence) demonstrate that LGA-Net achieves a mIoU of 97.3%, an OA of 99.0%, and a mAcc of 97.8%. These results indicate that the proposed method outperforms existing mainstream algorithms in both segmentation accuracy and robustness. This study not only proposes a feasible solution for precise sheep extraction under the tested experimental conditions, but also provides solid technical support for subsequent automated body measurement and behavior analysis.