G-Otsu：分層模型應用於同時分割多個影像

Abstract

影像分析是一種強大的方法能有效解決多種難題，能有效應用於農業領域。植物影像分析在農業中扮演著重要的角色，適用於植物研究以提高作物產量的研究效率和品質。本研究旨在探索基於植物特徵的演算法能夠快速進行非破壞性取樣 (non-destructive sampling) 進行外表型性狀分析。Otsu為單一通道的影像分割 (image segmentation) 方法，G-Otsu是一種多層次模型，結合了Otsu與多張影像之間的訊息，適合用於整個資料集分析。Otsu適用於雜訊與目標像素值差距大時，若雜訊貼近目標像素值，Otsu方法的影像分割的前景則容易涵蓋雜訊，而G-Otsu適用於解決此問題。G-Otsu透過不同摘要統計量的應用，本篇共有M、V、MV、K和L版本的G-Otsu方法。K和L版本G-Otsu使用了不同的係數調整，結合了影像之間的相似性與變異性進行影像分割。K版本的G-Otsu在擁有5~100張影像的資料集中，成功分割影像的機率較原始Otsu版本高。透過G-Otsu，我們能夠更好地利用影像之間的特徵訊息。未來，我們可以優化G-Otsu，藉由額外的植物特徵或使用多通道分析，結合多閾值進行影像分析，擴大G-Otsu於其他作物的溫室或生長室的應用。同時，也將結合其他影像分析方法和機器學習方法，以進一步提升分析的準確性。期望基於植物影像特徵的分析將在農業領域發揮重要的作用，為作物生產的提升與農業發展做出貢獻。

Image analysis is a powerful method with extensive applications in agriculture that can address various challenges effectively. Plant image analysis plays a crucial role in agriculture, enabling efficient and high-quality research based on plant phenotyping. The aim of this study is to explore algorithms based on plant features to conduct rapid non-destructive sampling for phenotypic trait analysis. The traditional Otsu algorithm is a single-channel image segmentation method. However, when noise pixel values are very close to the target pixel values, Otsu segmentation may inadvertently include noise. The G-Otsu algorithms are developed in this project, which is a multi-level model combining Otsu with additional information from multiple images. G-Otsu efficiently estimates multiple thresholds simultaneously for all images in the dataset. Five different versions of G-Otsu were tested and each employing different summary statistics: M, V, MV, K, and L versions. The K and L versions of G-Otsu employ distinct coefficient adjustments to incorporate both image similarity and variability into the segmentation algorithm. These two models show higher success rates in image segmentation compared to the original Otsu algorithm. In addition, the K-version of G-Otsu also demonstrates a higher success rate of image segmentation than Otsu in multiple datasets containing between 5 to 100 images. Through G-Otsu, we can better utilize feature information across images. In the future, we can further optimize G-Otsu by incorporating additional plant features or employing multi-channel analysis, integrating multiple thresholds for image segmentation, and expanding the application of G-Otsu to other crops in greenhouses or growth chambers. Additionally, combining other image analysis methods and machine learning techniques will further enhance the accuracy of analysis. It is expected that the analysis based on plant image features will play a crucial role in the agricultural field, contributing to crop production improvement and agricultural development.