Rice-SVBDete: A Detection Algorithm for Small Vascular Bundles in Rice Stem's Cross-Sections

XIAOYING ZHU^1*Weiyu Zhou¹Jianguo Li²Mingchong Yang²Haiyu Zhou²Jiada Huang²Jiahua Shi³Jun Shen⁴Pang Guangyao¹Lingqiang Wang^2*

• ¹Wuzhou University, Wuzhou, China
• ²Guangxi University, Nanning, Guangxi Zhuang Region, China
• ³Queensland Alliance for Agriculture and Food Innovation, University of Queensland, St Lucia, Queensland, Australia
• ⁴University of Wollongong, Wollongong, New South Wales, Australia

The vascular bundles in rice are crucial for its growth, development, and yield formation.Accurate measurement of their structure and distribution holds critical significance for rice breeding and cultivation. Although convolutional neural networks have achieved remarkable success in imagery object detection, detecting small vascular bundles from noisy backgrounds remains a great challenge due to their minuscule size. To address these challenges, this paper introduces "Rice-SVBDete", an algorithm specifically designed for detecting small vascular bundles in rice stem's cross-sections. First, the algorithm integrates Dynamic Snake-shaped Convolution (DSConv) into the Backbone network to enhance the detection of tiny objects.Second, we propose a Multi-scale Feature Fusion mechanism (MFF), which combines features at different scales from the Backbone network and the feature pyramid network and path aggregation network structures, thereby improving the model's ability to detect multi-scale targets. Finally, the Powerful Intersection over Union(PIoU) loss function replaces the Complete Intersection over Union(CIoU) loss function to focus more on the spatial consistency and positional accuracy between predicted and ground truth boxes, further improving small object detection's performance.The proposed method achieves a precision of 0.789, a recall rate of 0.771, and a mean average precision (mAP@.5) of 0.728 at IoU=0.50 for small vascular bundle detection. Compared with the original YOLOv8 network, where precision, recall, and mAP@.5 are improved by 0.179, 1 Sample et al.0.201, and 0.227, respectively. The proposed method provides valuable insights for enhancing the detection of small vascular bundles in rice stem's cross-sections.