A Multiscale Lightweight and Efficient Model Based on YOLOv7: Applied to Citrus Orchard

Area of Science:

• Agricultural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Human labor efficiency is a significant bottleneck in increasing citrus production.
• Developing unmanned citrus picking technology requires accurate, fast, and deployable detection models.
• Traditional object detection methods struggle to balance accuracy, speed, and model size.

Purpose of the Study:

• To develop an improved YOLOv7 model for efficient and accurate citrus detection.
• To address the limitations of traditional methods in achieving balanced performance for automated harvesting.
• To contribute to the development of unmanned citrus picking technology.

Main Methods:

• An improved YOLOv7 network was proposed, incorporating a small object detection layer.
• Lightweight convolution and a Convolutional Block Attention Module (CBAM) were integrated for multi-scale feature extraction.
• The model was optimized to reduce parameters and computational costs.

Main Results:

• The improved Citrus-YOLOv7 model achieved an average accuracy (mAP@0.5) of 97.29%.
• The average prediction time was 69.38 ms, demonstrating high efficiency.
• Model parameters and computation costs were reduced by 11.21 M and 28.71 G, respectively, compared to the original YOLOv7.

Conclusions:

• The Citrus-YOLOv7 model significantly improves citrus detection performance.
• The model offers a viable solution for automated citrus detection, addressing labor limitations.
• This research contributes to advancing unmanned harvesting technology in agriculture.