Rice-SVBDete: a detection algorithm for small vascular bundles in rice stem's cross-sections

Area of Science:

• Plant Science
• Agricultural Technology
• Computer Vision

Background:

• Vascular bundles are crucial for rice growth and yield.
• Accurate measurement of vascular bundles is essential for rice breeding.
• Detecting small vascular bundles in microscopy images is challenging due to size and background noise.

Purpose of the Study:

• To develop a specialized deep learning algorithm for detecting small vascular bundles in rice stem cross-sections.
• To improve upon existing object detection architectures for enhanced performance on small, intricate biological structures.

Main Methods:

• Proposed Rice-SVBDete algorithm, enhancing the YOLOv8 architecture.
• Incorporated Dynamic Snake-shaped Convolution (DSConv) for adaptive feature extraction.
• Implemented Multi-scale Feature Fusion (MFF) and a novel Powerful Intersection over Union (PIoU) loss function.

Main Results:

• Rice-SVBDete achieved a precision of 0.789, recall of 0.771, and mAP@.5 of 0.728.
• Demonstrated significant improvements over baseline YOLOv8, with precision, recall, and mAP@.5 increases of 0.179, 0.201, and 0.227, respectively.
• Effectively detected small vascular bundles in complex backgrounds.

Conclusions:

• Rice-SVBDete offers a robust solution for small object detection in challenging biological imaging.
• Provides a valuable tool for rice anatomical analysis, precision agriculture, and plant science research.
• Facilitates advancements in crop improvement and cultivation strategies.