Аннотация
This study introduces a novel YOLO-like multi-branch deep learning architecture designed for accurate apple detection and segmentation in orchard environments, addressing the persistent challenges of occlusion, illumination variability, and fruit clustering. The proposed model integrates an enhanced backbone with C2f modules and a Spatial Pyramid Pooling Fast (SPPF) block to capture multi-scale receptive fields, while a Feature Pyramid Network (FPN) combined with a Path Aggregation Network (PAN) ensures effective top-down and bottom-up feature fusion. To extend beyond bounding box localization, a prototype-based segmentation head is incorporated, enabling precise instance mask generation with reduced computational overhead. The model was comprehensively evaluated on the MinneApple dataset, consisting of high-resolution orchard images with polygonal annotations, and compared against state-of-the-art detection and segmentation frameworks, including Faster R-CNN, Mask R-CNN, SSD, YOLO variants, YOLACT, and SOLOv2. Quantitative results demonstrated that the proposed approach achieved superior mean Average Precision (mAP@0.5 = 0.76), precision (0.83), and F1-score (0.76), while maintaining a competitive inference speed of 40 FPS, confirming its suitability for real-time agricultural applications. Qualitative analysis further highlighted robustness in complex orchard conditions, reinforcing the model’s applicability for automated harvesting, yield estimation, and orchard monitoring. These findings advance the state of agricultural computer vision by unifying detection and segmentation in a lightweight, high-performance framework.