Real-Time Accurate Apple Detection Based on Improved YOLOv8n in Complex Natural Environments

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Efficient and accurate apple detection is critical for automated apple harvesting.
• Existing object detection models often face challenges with speed and accuracy, especially for occluded objects.

Purpose of the Study:

• To develop a lightweight and efficient apple detection model for robotic applications.
• To improve detection accuracy, speed, and robustness, particularly for partially occluded apples.

Main Methods:

• Proposed a novel lightweight apple detection model based on YOLOv8n.
• Introduced a Self-Calibrated Coordinate (SCC) attention module for enhanced feature extraction.
• Integrated a Partial Convolution Module improved with Reparameterization (PCMR) and Polynomial Loss (PolyLoss).
• Fused multi-scale features from the second and third pyramid levels for optimization.

Main Results:

• The improved model achieved an mAP of 88.90%, a 2.90% increase over YOLOv8n.
• Detection speed improved by 30.55% to 220 FPS, with reduced parameters and FLOPs.
• Outperformed mainstream object detection algorithms in both mAP and speed.
• Achieved 40 FPS on an Android application, a 26.93% improvement over YOLOv8n.

Conclusions:

• The proposed lightweight model significantly enhances apple detection efficiency and accuracy for robotic systems.
• The model's performance and real-time capabilities make it suitable for resource-constrained environments.
• This research offers a valuable reference for developing efficient detection models in agricultural robotics.