Plant health monitoring and fruit counting are essential for optimizing crop yields, minimizing economic losses, and producing high-quality agricultural products. This study focused on identifying an optimal deep learning model for real-time and accurate automated visual assessment of orange health and yield, specifically targeting deployment on mobile devices with limited resources. The performance of four lightweight YOLO (You Only Look Once) models is compared, for the real-time detection and classification of oranges (Yolov5nu, Yolov9t, Yolov10n, and Yolo11n). The evaluation of YOLO models considered precision, recall, mAP50, mAP50-90, and inference time. YOLOv10n achieved the highest precision (88.8%), YOLOv9t combined strong recall (80.5%) with the best mAP50 (87.6%), while YOLOv5nu was the fastest (1.6ms). Overall, YOLO11n provided the best balance of precision (87.2%), recall (80.5%), and low inference time (1.7ms), making it ideal for real-time use. As a result, YOLO11n was deployed in an Android app for on-device orange detection, classification, and counting, with real-time CSV logging. This demonstrates that optimized YOLO models can deliver efficient, accurate, and practical solutions for mobile-based smart farming.