Rice Grain Detection and Counting Method Based on TCLE-YOLO Model.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate rice yield estimation relies on thousand-grain weight, necessitating precise rice grain detection and counting.
• Challenges include small grain size, high similarity, and adhesion, hindering traditional measurement methods.

Purpose of the Study:

• To develop an advanced deep learning model for accurate rice grain detection and counting.
• To improve the reliability of thousand-grain weight measurements for rice breeding and cultivation.

Main Methods:

• A TCLE-YOLO model was designed using YOLOv5 as the backbone, incorporating a coordinate attention (CA) module for enhanced small target feature representation.
• A specialized detection head for small targets was added, utilizing low-level, high-resolution feature maps.
• A transformer encoder was integrated into the neck module to expand the receptive field and improve feature extraction, particularly for adhesive grains.
• EIoU loss was employed to further refine detection accuracy.

Main Results:

• The TCLE-YOLO model achieved high performance on a self-built dataset, with precision, recall, and mAP@0.5 reaching 99.20%, 99.10%, and 99.20%, respectively.
• The model demonstrated superior detection performance compared to several state-of-the-art models.
• The enhanced sensitivity to small and adhesive grains was confirmed.

Conclusions:

• The TCLE-YOLO model provides an effective solution for rice grain recognition and counting.
• This method offers valuable support for accurate thousand-grain weight measurements and efficient rice breeding evaluations.