Related Experiment Video
Updated: Jul 6, 2025
11:49
Cereal Crop Ear Counting in Field Conditions Using Zenithal RGB Images
Published on: February 2, 2019
9.3K
A lightweight network for improving wheat ears detection and counting based on YOLOv5s
Xiaojun Shen1, Chu Zhang1, Kai Liu1
1School of Information Engineering, Huzhou University, Huzhou, China.
Frontiers in Plant Science
|January 2, 2024
Summary
This study introduces a lightweight deep learning model for real-time wheat ear detection and counting. The optimized YOLOv5s model achieves high accuracy with reduced computational resources, enabling precision agriculture on limited hardware.
Area of Science:
- Agricultural Science
- Computer Vision
- Deep Learning
Background:
- Accurate wheat ear recognition is vital for yield prediction, but current deep learning methods are computationally intensive for real-time field applications.
- Existing models face challenges like large parameter volumes, heavy model weights, and slow processing speeds, hindering deployment on resource-constrained devices.
- There is a significant need for lightweight wheat ear detection methods suitable for real-time identification in wheat fields.
Purpose of the Study:
- To develop a lightweight and efficient deep learning model for real-time wheat ear detection and counting.
- To optimize the YOLOv5s model for reduced computational complexity and improved performance on limited hardware.
- To enhance the model's accuracy and robustness for diverse wheat ear variations and environmental conditions.
Main Methods:
- Proposed a lightweight wheat ear detection method based on YOLOv5s, optimizing it with the ShuffleNetV2 network to decrease parameters and computational complexity.
- Introduced a content-aware reassembly of features (CARAFE) as a lightweight upsampling operator in the feature pyramid to maintain detection performance.
- Incorporated a dynamic target detection head to adaptively adjust detection strategies for variations in wheat ear scale, shape, and orientation.
Main Results:
- The optimized model achieved a weight of 2.9 MB, 2.5 * 10^9 FLOPs, and a mean Average Precision (mAP) of 94.8%.
- Linear fitting determination coefficients (R^2) were 0.94 for the global dataset and 0.97 for the local experimental site, indicating strong agreement between predicted and actual counts.
- The model demonstrated robustness and generalization capabilities when tested on both global and local datasets.
Conclusions:
- The developed lightweight model effectively meets the demands for precise wheat ear counting in real-time.
- The optimized approach significantly reduces computational load, making it suitable for embedded systems and mobile devices in precision agriculture.
- This research contributes to advancing real-time agricultural monitoring and yield prediction technologies.
