基于改进YOLO v8n的多育种场景下大豆主茎表型解析方法

Soybean Main Stem Phenotype Analysis Method Based on Improved YOLO v8n in Multiple Breeding Scenarios

摘要:

主茎表型准确解析对理解植株生长模式、优化种植密度以及提高作物产量具有重要意义。本文提出了一种基于改进YOLO v8n的多育种场景下大豆主茎节点检测模型（YOLO?SN），并对主茎节点数量、节间距和主茎长度表型参数进行了解析。在主干构建DAF模块（Deformable attention fusion module），增强模型对多场景下节点区域关注能力;在头部设计动态解耦头（Dynamic decouple?head，DDH），提升模型对大豆主茎节的感知能力;引入动态非单调聚焦机制（Wise intersection over union，WIoU），提升网络收敛速度与模型泛化能力。试验结果表明，在单一场景下主茎节检测准确率、召回率、平均精度均值和F1分数均有明显提升，在多场景下综合性能分别达90.6%、85.1%、90.0%和87.8%，均优于主流目标检测模型。主茎节点数量、节间距和主茎长度的绝对误差、决定系数分别为0.42、12.6像素、17.4像素和0.88、0.80、0.82，实现了多场景下大豆主茎表型高精度解析。研究结果为不同育种环境下大豆植株表型准确解析提供了有效方法。

Abstract:

Parsing of main stem phenotype is of great significance for understanding plant growth patterns, optimizing planting density and improving crop yield. In this study, one soybean main stem node detection model YOLO?SN based on improved YOLO v8n was proposed for multiple breeding scenarios, and the phenotypic parameters of main stem node number, internode length and main stem length were analyzed. Firstly, aiming at the problems of poor model adaptability and high missed detection rate caused by significant differences in the characteristics of soybean main stem nodes in three typical breeding scenarios, including open field, indoor and artificial climate chamber, respectively. The DAF (Deformable attention fusion module) module is constructed in the backbone to enhance the model?s ability to pay attention to node regions in multiple scenarios;Secondly, for the problem that the small pixel share of main stem nodes leads to recognition difficulties, the DDH (Dynamic decouple?head) was designed in the head to improve the model?s ability to perceive the main stem nodes of soybeans;Finally, the dynamic non?monotonic focusing mechanism WIoU (Wise intersection over union, WIoU) was designed for further improving the convergence speed and generalization of original network YOLO v8n. The experimental results show that the main stem node detection accuracy, recall, average precision mean and F1?score are significantly improved in a single scene, and the comprehensive performance in multiple scenes reaches 90.6%, 85.1%, 90.0% and 87.8%, respectively, which are all better than the mainstream target detection models. The absolute error and coefficient of determination of the number of main stem nodes, the distance between nodes and the length of main stem were 0.42, 12.6 pixels, 17.4 pixels and 0.88, 0.80, 0.82, respectively. This study provide an effective method for the accurate parsing of soybean plant phenotypes in typical breeding environments, as well as offering technical support for intelligent breeding of soybeans and other legumes.