Visual weight estimation of caged laying hens via semantic segmentation and ensemble regression

• Alightweight multi-stage framework enables accurate, non-contact weight estimation of caged laying hens. • Cross-view ensemble learning and EnhancedSegFormer improve robust ness under complex farming conditions. Accurate non-contact methods of estimating the body weight of laying hens are essential for intelligent poultry farming. This study proposes a multi-stage visual analysis framework that is specifically designed for caged environments, and which addresses key challenges such as occlusion, posture variation, background complexity, and variable imaging distances. The framework consists of three core modules: (1) a lightweight Transformer-based segmentation network called EnhancedSegFormer, which incorporates cross-scale attention and adaptive feature fusion to achieve robust segmentation across varying distances (mIoU: 89.6%, BF1: 88.5%); (2) a geometric feature-driven random forest classifier for automated orientation recognition (side vs. back view), which attains F1-scores that are consistently above 91% across the 40–100 cm range; and (3) a stacking ensemble regression model that integrates XGBoost and TabTransformer to predict body weight based on multi-view morphological features. The ensemble model yields the highest accuracy at distances of 40–59 cm, with an RMSE of 0.088 kg and R² of 0.944. Experimental results validate the proposed strong generalisation, stability, and deployment readiness of the framework for real-world farming conditions. This approach provides a practical, real-time, and stress-free solution for precision weight monitoring in laying hens, contributing to enhanced data-driven livestock management.