Hierarchical Segmentation Of Leaf Venation Based On Region Growth

Leaf veins are important structures that support leaf growth and transport the nutrients required by the leaf and photosynthetic products.The leaf veins can be classified according to their location,growth trend,width and other factors.For example,a primary vein is defined as the thickest leaf vein extending from the petiole to the tip of the leaf.The leaf vein hierarchy can better characterize the leaf venation compared to the leaf vein network.Hierarchical segmentation of leaf veins is of great importance image classification,leaf modeling and molecular breeding.However,most of the current research on leaf veins is confined to the leaf vein network.Due to the complexity of the venation,the hierarchy of veins that can better explore the properties of leaf veins is rarely involved.On the one hand,leaf veins at all hierarchies have high similarity in color and brightness,which are challenging to be segmented directly by threshold.On the other hand,leaf veins have variable trends,and a few methods using unidirectional linear filtering are not comprehensive enough to segment leaf veins.To address these difficulties,this thesis proposed a region-growth-based leaf vein hierarchical segmentation algorithm and designed a fully automated leaf vein phenotype extraction process.The main research contents are as follows.(1)An algorithm was designed for leaf vein hierarchical enhancement segmentation based on multi-scale morphological filtering eigenvalues.The problem of leaf veins hierarchy is hard to be segmented directly by threshold.The method enhanced the contrast of leaf vein based on the eigenvalues of morphological filtering response,and introduced the multiscale idea for enhancing leaf veins at all hierarchies according to the leaf vein hierarchy characteristics.Quantitative experiments were designed to evaluate the effectiveness of the method.The experimental results displayed that the enhanced leaf vein is evidently better than the unenhanced,which was validated in cherry and soybean leaf datasets,proving the validity of the method.(2)A region-growing algorithm based on leaf vein orientation information was designed for hierarchical leaf vein reconstruction to solve incomplete leaf vein hierarchical segmentation by linear filtering.The method obtained the eigenvector of pixels in the curve-like leaf vein by sorting the morphological filtering response values of leaf vein pixels,and used them as directional information to design a leaf vein region growth algorithm which improved the effect of leaf vein segmentation through post-processing.Hierarchical segmentation of veins was conducted on soybean and cherry datasets.The results expressed that the method outperformed the existing methods in the segmentation of primary and secondary leaf veins.The average completeness of primary leaf veins in the cherry dataset reached 0.87,with average deviation completeness less than 2 pixels.The average accuracy of secondary leaf veins reached 0.82,with average deviation correctness less than 5pixels.On this basis,the method attempts to segment of tertiary leaf veins,which to our knowledge,is the first segmentation for tertiary leaf veins.(3)To address the problem of inadequate leaf vein phenotype extraction,a fully automated phenotype extraction process was designed based on the results of leaf vein hierarchical segmentation.The process of leaf vein phenotype extraction was described in detail,including image scanning and labeling,pre-processing,leaf vein hierarchical segmentation,and leaf vein phenotype extraction base on network model.Based on the extraction results of this thesis,the existing leaf vein phenotypes are supplemented and improved.The network model of leaf veins is introduced in the phenotype extraction and simplified by the Douglas-Puke algorithm,which improved the efficiency of the phenotype extraction algorithm.Leaf angle was quantitatively analyzed on soybean as an example,and reached an average accuracy of 95.79%,indicating the dependability of the network model.The main contribution of this thesis is to propose solutions to the existing problems in leaf vein hierarchical segmentation.This thesis designed the corresponding algorithms with certain generality in dicotyledons,and further improved the leaf vein phenotype database,to provide data support for the subsequent research in agronomy and biology.