Precision Detection of Dense Plums in Orchards Using the Improved YOLOv4 Model

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Precision detection of dense, small targets like plums is crucial for agricultural robots.
• Existing visual detection algorithms struggle with plum recognition due to size and dense growth patterns.

Purpose of the Study:

• To develop a lightweight and accurate model for detecting dense plums in orchards.
• To improve the performance of visual perception systems for agricultural picking robots.

Main Methods:

• An improved YOLOv4 (You Only Look Once version 4) model was proposed, utilizing MobilenetV3 as the backbone and depthwise separable convolution for efficiency.
• A 152x152 feature layer was introduced for fine-grained detection of dense targets.
• Category balance data augmentation, Focal loss, and complete intersection over union (CIOU) loss were employed.

Main Results:

• The improved YOLOv4 model achieved superior mean average precision (mAP) compared to YOLOv4, YOLOv4-tiny, and MobileNet-SSD.
• Model size was reduced by 77.85%, parameters by 82.08%, and detection speed increased by 112% relative to YOLOv4.
• The model demonstrated strong robustness and high accuracy under varying illumination, intensity, and occlusion conditions.

Conclusions:

• The proposed lightweight YOLOv4 model offers a significant advancement in dense plum detection for agricultural applications.
• The model's efficiency and robustness make it suitable for real-world orchard environments, supporting yield estimation and robotic harvesting.