Research on Sheep Mounting Behavior Recognition in Complex Scenes Based on an Improved YOLOv11

Mounting behavior in sheep is a critical ethological indicator for identifying the estrus status of ewes, and plays an essential role in breeding management and estrus monitoring. Traditional manual observation methods suffer from subjectivity, low efficiency, and high omission rates, which limit their applicability in large-scale intelligent farming. To address the challenges of accurately recognizing mounting behavior in complex farm environments, such as dramatic illumination changes, severe occlusions, and dense sheep clusters, this study aimed to develop a high-precision and robust automatic recognition model to enable rapid detection and precise localization of mounting behavior, thereby supporting intelligent sheep reproduction management.

Daily activity videos of 24 ewes were collected from the Jiaxiang Breeding Sheep Farm in Shandong Province between April 15 and May 15, 2024. A balanced dataset consisting of 4 700 annotated images (including both mounting and non-mounting samples) was constructed. Based on the YOLOv11 architecture, an improved detection model was proposed, named SIDS-YOLOv11, which incorporated four key modules: SCINet for low-light image enhancement, improving visual quality in dim conditions; iAFF for optimizing multi-scale semantic feature fusion; DySample for enhancing edge detail recovery via dynamic upsampling; and SEAM for improving target perception under occlusions using spatial attention. The training process employed the CIoU (Complete Intersection over Union) loss function for bounding box regression, combined with various data augmentation techniques to enhance model robustness and generalization.

On the validation set, compared with the original YOLOv11 model, SIDS-YOLOv11 achieved a mAP@0.5 of 0.942, a Precision of 0.956, a Recall of 0.854, and a mAP@0.5-0.95 of 0.703—representing improvements of 3.5%, 4.7%, 1.7%, and 1.5%, respectively. Heatmap visualizations demonstrated that the improved model maintained accurate focus on target regions even in low-light and occluded scenarios. The attention regions of the enhanced model were more concentrated, background noise was reduced, and feature extraction capabilities were significantly improved, leading to enhanced recognition accuracy and localization stability. Compared with mainstream detection models, such as YOLOv8n, YOLOv6, Faster R-CNN, and SSD, SIDS-YOLOv11 achieved a better balance between detection accuracy and inference speed. Evaluation on low-light and heavily occluded videos further verified the model's superior performance, maintaining over 63% detection accuracy in complex scenarios, indicating strong applicability and adaptability.

The proposed SIDS-YOLOv11 model effectively integrated image enhancement, feature representation, and attention mechanisms, significantly improving the recognition accuracy of sheep mounting behavior in complex environments. The model maintained stable performance under challenging conditions, such as occlusion and low illumination, offering a high-performance visual recognition solution for estrus monitoring, behavior analysis, and breeding management in smart farming. The model held strong potential for practical deployment and large-scale application.