Heterogeneous models with synergistic feature fusion for real-time apple recognition in robust apple harvesting robotics

The advancement of agricultural automation intensifies demands for highly reliable visual recognition systems in orchard harvesting robots. However, it is difficult to achieve robust fruit detection for balancing high accuracy and real-time performance under complex orchard conditions with variable illumination, occlusions, and phenotypic diversity. This study proposes a collaborative framework integrating adaptive image processing with heterogeneous model inference. The methodology begins with Lab color space conversion to ensure illumination invariance. It further utilizes dual-threshold HSV segmentation for handling both red and green apples, alongside morphological optimization with elliptical structuring elements to address occlusion. A novel architecture allocates real-time screening to an embedded Random Forest (RF) classifier and precise localization to a host-based lightweight YOLOv5 model through fused color-morphological features. Experimental results demonstrate that morphological feature enhancement consistently outperforms color-based approaches across both models. The models with dual-feature input achieves optimal performance that the RF classifier attains accuracy of 83.35%, while lightweight YOLOv5 reaches 98.90% accuracy. Quantitative analysis reveals dual-feature fusion with color and morphology improving all metrics by over 3% compared to non-enhanced baselines. Notably, the observed accuracy improvement exceeded the sum of gains from individual features, confirming a synergistic effect and proving the necessity of feature fusion. This work provides a computationally viable solution for reliable apple recognition in unstructured environments.