IEEE/ACM Trans Comput Biol Bioinform. 2024 May-Jun;21(3):372-382. doi: 10.1109/TCBB.2024.3364208. Epub 2024 Jun 5.
Plant stomatal phenotype traits play an important role in improving crop water use efficiency, stress resistance and yield. However, at present, the acquisition of phenotype traits mainly relies on manual measurement, which is time-consuming and laborious. In order to obtain high-throughput stomatal phenotype traits, we proposed a real-time recognition network SLPA-Net for stomata localization and phenotypic analysis. After locating and identifying stomatal density data, ellipse fitting is used to automatically obtain phenotype data such as apertures. Aiming at the problems of small stomata and high similarity to background, we introduced ECANet to improve the accuracy of stoma and aperture location. In order to effectively alleviate the unbalance problem in bounding box regression, we replaced the Loss function with a more effective Focal EIoU Loss. The experimental results show that SLPA-Net has excellent performance in the migration generalization and robustness of stomata and apertures detection and identification, as well as the correlation between stomata phenotype data obtained and artificial data.
植物气孔表型特征在提高作物水分利用效率、抗逆性和产量方面发挥着重要作用。然而，目前表型特征的获取主要依赖于人工测量，既耗时又费力。为了获得高通量的气孔表型特征，我们提出了一种用于气孔定位和表型分析的实时识别网络 SLPA-Net。在定位和识别气孔密度数据后，使用椭圆拟合自动获取孔径等表型数据。针对小气孔和与背景相似度高的问题，我们引入了 ECANet 来提高气孔和孔径定位的准确性。为了有效缓解边界框回归中的不平衡问题，我们用更有效的 Focal EIoU Loss 替换了 Loss 函数。实验结果表明，SLPA-Net 在气孔和孔径检测和识别的迁移泛化和鲁棒性方面，以及与人工数据获得的气孔表型数据之间的相关性方面表现出优异的性能。
