ECL-YOLOv8: An improved YOLOv8-based algorithm for detecting cattle faces and muzzles in complex open pasture environments

Abstract Background An efficient and reliable method for detecting cattle faces and muzzles supports automated livestock management and intelligent detection.Both the face and muzzle of cattle contain visual information that exhibits significant inter-individual variation. Computer vision-based non-contact identification methods offer advantages such as high efficiency, low cost, and animal welfare friendliness. However, much of the current research is limited to close-up shots or manually selected images of cattle faces or muzzles, with little consideration given to identifying facial or nasal locations in the complex environment of open pastures. This study aims to evaluate the model’s robustness in complex environments by conducting multi-angle imaging under various conditions, including varying distances, obstructions from livestock facilities, cattle grazing, muddy faces, and exposure variations. Therefore, this study uses YOLOv8 as the foundation to design the ECL-YOLOv8 model for the simultaneous detection of cattle faces and muzzles in complex open pasture environments. Results Experiments conducted on our in-house dataset show that the ECL-YOLOv8 model achieved an accuracy of 92.6%, a recall of 81.7%, and a mean precision of 87.4%. Compared to YOLOv8, ECL-YOLOv8 achieved improvements of 1, 4.2, and 3.1 percentage points in accuracy, recall, and mean precision, respectively. Specifically, the mean precision for muzzle detection was 82.4%, which is 5.3 percentage points higher than the baseline. These results demonstrate that ECL-YOLOv8 exhibits greater robustness under the complex conditions found on ranches. Conclusions This study investigates the problem of accurately recognizing cattle faces and muzzles in the complex environment of open pastures and proposes a detection model—ECL-YOLOv8—based on an improved YOLOv8 architecture. The model effectively enhances the detection capabilities of cattle faces and muzzles in complex pasture environments. In particular, the model demonstrates significant advantages in muzzle detection performance and adaptability to complex backgrounds. Future work will focus on improving inference efficiency and model lightweighting to further support practical deployment, thereby better serving real-time monitoring and intelligent operations in actual pastures and providing more reliable technical support for smart agriculture applications.