Real-time detection method for Litchi diseases and pests based on improved YOLOv5s

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Traditional manual detection of Litchi pests and diseases is inefficient and costly.
• Sustainable orchard management requires accurate, real-time monitoring solutions.
• Existing deep learning models may be too computationally intensive for embedded platforms.

Purpose of the Study:

• To develop an improved YOLOv5s model for efficient, real-time Litchi pest and disease detection.
• To optimize the model for deployment on resource-constrained embedded platforms.
• To enhance detection accuracy and reduce computational overhead.

Main Methods:

• Modified YOLOv5s backbone with ShuffleNetV2 for reduced parameters and computation.
• Integrated depthwise convolutions (DWConv) and C3Ghost modules in the detection head.
• Incorporated attention mechanisms (SE, CBAM, CoordAtt) and EIoU loss function.

Main Results:

• Achieved a mean average precision (mAP) of 96.7% for Litchi pest and disease detection.
• Reduced computational cost by 87.5%, parameters by 86.7%, and model size by 55.6% compared to original YOLOv5s.
• Attained an inference speed of 3.3 FPS on Raspberry Pi 4B, a 57.1% improvement, meeting real-time requirements.

Conclusions:

• The YOLOv5s-SNV2-GSE model provides a practical and efficient solution for real-time Litchi pest and disease detection.
• The model's optimization makes it suitable for deployment in resource-constrained environments for sustainable agriculture.
• Significant reductions in computational cost and model size enable wider application of advanced AI in orchards.