Research On Leaf Phenotypic Traits Extraction Based On Vein Segmentation

With the combination of computer vision technology and intelligent agriculture management,the acquisition of plant leaf phenotypic data becomes more convenient.As an important tissue in plant leaves,the quantifying analysis of vein network structure is of great significance to leaf phenotypic parameters extraction.However,current studies on this field have not truly achieved high-throughput automated parameter extraction.Specifically,on the one hand,most of current researches still rely on vein bookmark images,which requires manual processing,to acquire vein network structures.For leaf scanning image,affected by the imaging environment,illumination and other factors,the segmentation results need to be improved.On the other hand,due to the complexity of vein network structure,few studies have made attempts to automatically divide vein levels.Leading to the fact that the extraction of meaningful vein hierarchical structure traits still requires manual measurement.Thus,raising a new challenge yet to be solved.Aimed at the two major problems on plant leaf phenotype extraction,the main contributions of this thesis are listed below:For the self-acquired leaf scanning images,an adaptive morphological enhancement method based on Hessian matrix is proposed for vein network segmentation.The necessity of linear structure enhancement is verified through comparing and analyzing the effect of existing vein segmentation methods on leaf scanning images.With the linearity of veins,multi-scale Hessian matrix is used to calculate the direction information of each pixel.Through the analysis of internal structure,pixels located in linear structure are searched for selective dilation or erosion along their main direction.Thus,an adaptive morphological enhancement method based on Hessian matrix is proposed.Experimental results show that the proposed method can enhance linear structure,while suppressing noise to improve the quality of derived vein network images.Meanwhile,achieving better segmentation results on both leaf scanning images and vein bookmark images.Aimed at the problem that existing researches on phenotype extraction can not automatically calculate vein hierarchical structure traits,a method of leaf vein Hierarchy segmentation based on Two Step Region-based Growing is proposed.With the division of vein regions through region grow,the basic unit for analysis is promoted from single pixel to vein segment region.Further,the main orientation of each vein level is determined using the topological response.Meanwhile,Due to the presence of noise in binary vein network images,a robust directional filter with hit or miss constraints is designed to locate initial vein segments of each level.Finally,a vein reconstruction strategy based on segment-linking is proposed to iteratively locate vein segment regions belonging to the same vein branch.To verify the performance of proposed method,a soybean transmission image dataset is collected for the lack of publicly available leaf dataset with vein hierarchy marks.And the first and second order veins are roughly labelled by experts according to the proposed vein hierarchy model.The method is explored through three aspects,including accuracy,invariance to translation,rotation and scaling,and the effectiveness of segment-linking strategy.The experimental results show that the proposed method works pretty well while improving the performance in terms of completeness and with no compromise on correctness.The average deviation on major vein is less than 5 pixels and the average completeness on second-order veins reaches 54.28%.Furthermore,it is proved to be invariance to translation,rotation and scaling.Taking soybean leaves as research object,a soybean phenotype extraction scheme is proposed with the designed vein segmentation method.The construction mainly consists of four steps: leaf image data acquisition and preprocessing;vein network segmentation;vein hierarchy segmentation;and leaf phenotype extraction.An iterative thinning algorithm is used to extract the skeleton of vein network images,and the nodes,tips and edges are defined according to the definition of graph for the quantifying description of vein network.Taking three typical soybean leaf images for example,four phenotypic parameters including edge density of first and second order veins,their branching angle,length and density of the second order vein branches are extracted for qualitative analysis.And quantitative analysis is performed on the calculation of branching angle.The average accuracy rate is 94.58%,which has a high reliability and provides certain reference for guiding the targeted breeding of soybean crops.The main contribution of this thesis is to describe the key difficulties in the research of plant leaf phenotype extraction,and to propose corresponding vein segmentation methods for soybean leaves under the specific demand scenarios.This can greatly promote the research on plant phenotype extraction,and provide early technical support for the forthcoming plant phenotype analysis and crop designed breeding.