A Lightweight Apple Detection Method in Real Orchard Environments Based on Improved YOLO

Optimizing object detection in orchard environments presents significant challenges due to variations in lighting, occlusions, and fruit positioning. This paper proposes a lightweight apple detection framework based on an improved YOLO architecture, integrating YOLOv12 with Soft-NMS, Weighted Box Fusion (WBF), and Grad-CAM explainability. The proposed system effectively localizes and classifies apples at different ripening stages—ripe, unripe, and overripe—while reducing false detections and enhancing interpretability through visual explanation maps. Experimental evaluations demonstrate that the combination of advanced non-maximum suppression strategies and ensemble fusion techniques enhances detection robustness and reliability. The framework is complemented by a Python-based interface for streamlined dataset processing and prediction, illustrating a practical deployment pathway for orchard monitoring applications.

Keywords: Apple detection, YOLOv12, Soft-NMS, Weighted Box Fusion, Grad-CAM, orchard monitoring, object detection, computer vision, deep learning, image classification.