Lightweight Algorithm for Apple Detection Based on an Improved YOLOv5 Model.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Apple-picking robots require efficient detection algorithms for unstructured environments.
• Existing algorithms often have large model sizes and slow inference speeds, limiting their use on embedded platforms.

Purpose of the Study:

• To propose a lightweight and efficient apple detection model for robotic harvesting.
• To enhance the performance of apple detection algorithms for embedded systems.

Main Methods:

• Introduced a lightweight C3-light module to replace C3 for improved spatial feature extraction and speed.
• Incorporated the SimAM attention module into the neck layer to boost model accuracy.
• Developed the YOLOv5-CS model based on YOLOv5n for apple detection.

Main Results:

• The YOLOv5-CS model achieved a size of 6.25 MB and an inference speed of 0.014 s.
• Reduced floating-point operations (FLOPs) by 15.56% compared to the baseline.
• Attained an average precision (AP) of 99.1%, outperforming mainstream networks.

Conclusions:

• The YOLOv5-CS model offers an efficient solution for real-time apple detection in complex orchard settings.
• This lightweight model provides technical support for intelligent apple-picking devices and visual recognition systems.
• The proposed model demonstrates superior performance in terms of accuracy, speed, and model size for robotic applications.