Lightweight fruit detection algorithms for low-power computing devices

A lightweight fruit detection algorithm is important to ensure real-time detection on low-power computing devices while maintaining detection accuracy. In addition, the fruit detection algorithm is also faced with some environmental factors. To solve these challenges, lightweight detection algorithms termed YOLO-Lite, YOLO-Liter and YOLO-Litest were developed based on the YOLOv5 framework. The compared mean average precision (mAP) detection revealed that YOLO-Lite at 0.86 is 2%, 4%, 5%, 7%, and 16% more than YOLO-Liter and YOLOv5n at 0.84 each, YOLOv4-tiny at 0.82, YOLO-Liter at 0.81, YOLO-MobileNet at 0.79, and YOLO-ShuffleNet at 0.70, respectively, but not for YOLOv8n at 0.87. On the Computer platform, except for YOLOv4-tiny at 178.6 frames per second (FPS), the speed of YOLO-Litest at 158.7 FPS is faster than YOLO-Liter at 129.9 FPS, YOLO-Lite at 120.5 FPS, YOLO-ShuffleNet at 119.0 FPS, YOLOv8n at 116 FPS, YOLOv5n at 111.1 FPS, and YOLO-MobileNet at 89.3 FPS. Using Jetson Nano, the 32.3 FPS of YOLO-Litest is faster than other algorithms, but not YOLOv4-tiny's 34.1 FPS. On the Raspberry Pi 4B, YOLO-Litest with 4.69 FPS, outperformed other algorithms. The choices for an accurate and faster detection algorithm are YOLO-Lite and YOLO-Litest respectively, while YOLO-Liter maintains a balance between them.