Segmentation and stacking ensemble learning-based method for estimating residual feed in rabbit cage feeders

Precision feeding is often required to monitor the residual feed levels in standardized meat rabbit farming. However, manual inspection cannot fully meet the intensive production in recent years, due to the labor-intensive and human error. While emerging 3D vision monitoring can also remain expensive and fragile under harsh breeding environments. Furthermore, conventional 2D deep learning failed to balance inference speed with regression accuracy on resource-constrained edge devices, particularly for the strong nonlinearity between 2D projected features and 3D feed weight. In this study, a robust, low-cost, and real-time system of residual feed estimation was developed to deploy on an autonomous inspection robot. Two-stage architecture with the decoupled "Segmentation-regression" was also designed to balance between visual perception complexity and limited edge computing power. In the first stage, the YOLOv8-seg instance segmentation model was employed to rapidly and accurately extract the feed region from the complex background of the rabbit cage. Advanced single-stage anchor-free architecture was utilized in the C2f module to optimize gradient flow, while a decoupled head structure was to separate classification from localization. Thereby the high-precision segmentation was obtained for the irregular feed boundaries without the computational overhead of two-stage models like Mask R-CNN. In the second stage, the four geometric features—projection area (S), contour perimeter (Z), and the length (L) and width (W) of the minimum bounding rectangle—were extracted rather than only on pixel area. A multidimensional feature vector was then constructed using these geometric features. The physical constraints were used to address the ambiguity in 2D-to-3D mapping caused by the inclined side walls of the feeders and the irregular accumulation resulting from rabbit foraging behaviors. Stacking ensemble regression model was constructed to map these features to weight. Three heterogeneous base learners were integrated: eXtreme Gradient Boosting (XGBoost), Random Forest (RF), and Back Propagation Neural Network (BPNN). A linear regression meta-learner was utilized to combine the predictions of these base models. The tree models were effectively balanced for tabular data to avoid the overfitting risks with single neural networks. The improved model was deployed on an embedded NVIDIA Jetson Xavier NX platform after TensorRT optimization. A trade-off analysis between FP32 and INT8 quantization mode was conducted for geometric fidelity. Experimental results indicated that the exceptional performance was achieved on the constructed dataset. The YOLOv8-seg model was attained a mean Average Precision (AP50) of 0.995 and a Mean Pixel Accuracy (MPA) of 0.9695, indicating the fine-grained edge features after extraction. In the regression task, the Stacking ensemble model was achieved in a Mean Absolute Error (MAE) of 2.3458 g and a coefficient of determination (R²) of 0.9904. A internal baseline comparison revealed that the Stacking strategy was reduced the MAE by 53.2% compared with a single BPNN model, indicating the ensemble strategy in the small-sample geometric regression. The deployment tests showed that the "sawtooth" edge noise was introduced to degrade the regression accuracy, while INT8 quantization offered the higher speeds. Conversely, the FP32 precision mode also achieved an inference speed of 29.2 frames per second (FPS). Real-time detection was achieved, meeting the 10 FPS requirement for the inspection robot traveling at 0.2 m/s. In conclusion, a lightweight and decoupled machine vision framework was validated to effectively treat the ambiguity of 2D features using feature engineering and ensemble learning. The high precision, strong robustness, and low hardware costs can offer the practically technical solution for the precision feeding in meat rabbit farming.