Rice Ear Counting Based on Image Segmentation and Establishment of a Dataset

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Manual rice ear counting for yield estimation is time-consuming, inefficient, and subjective.
• Computer vision offers a promising alternative to improve accuracy and efficiency in field assessments.

Purpose of the Study:

• To develop an automated system for real-time rice ear detection and counting in complex field conditions.
• To establish a comprehensive dataset for training and evaluating rice ear counting models.

Main Methods:

• Creation of a 3300-sample rice ear dataset with diverse conditions (lighting, background, occlusion).
• Application of data augmentation techniques to expand the dataset size.
• Development of a hybrid model combining a localization-based counting fully convolutional neural network (LC-FCN) with the watershed algorithm, utilizing transfer learning.

Main Results:

• The proposed LC-FCN and watershed algorithm model outperforms traditional machine learning and SSD algorithms for rice ear detection.
• Achieved a Mean Absolute Error (MAE) of 2.99 on a 300-sample test set.
• The model demonstrates advanced and innovative capabilities for dense rice image recognition.

Conclusions:

• The developed computer vision model provides an accurate and efficient method for counting rice ears in fields.
• This technology can supply crucial data for rice yield estimation and contribute to agricultural research.
• The established dataset serves as a valuable resource for future studies in rice crop monitoring.