Research On The Method For Plant Leaves Stomatal Segmentation Based On Mask R-CNN

Plant stomata are special structure on the leaf epidermis.They are the gate for plant individuals to exchange material and energy with the outside world.They play an extremely significant role in the interaction between plants and the environment,and the carbon-water cycle of the ecosystem.The morphological characteristics(density,size,etc.)and behavior(the size and frequency of stomatal pore opening)of plant leaf stomata affect the flux of carbon dioxide and water vapor,which in turn affects the photosynthesis and carbon sequestration capacity of plants.Accurate measurement of stomatal pore anatomical parameters is the key to the analysis of stomatal morphological characteristics and behavior.At present,the analysis of plant leaf stomatal pore anatomical parameters mainly adopts manual measurement or semi-automatic analysis methods assisted by the existing software,which is difficult to achieve rapid,accurate and high-throughput measurement.In response to this problem,this paper studies the automatic segmentation of plant leaf stomatal pore based on convolutional neural network.The main research contents are as follows:Image acquisition of living populus nigra leaves with ultra-depth-of-field optical microscope,high-resolution stomatal microscopic images obtained by deep synthesis technology;American populus balsamifera and Ginkgo leaf microscopic datasets were obtained from public data;Labelme labeling tool is used to manually mark the stomatal pores from the leaf microscopic images,extract the stomatal pores mask,and make the stomatal pores datasets for segmentation.The stomatal pore segmentation method based on convolutional neural network Mask R-CNN was studied and proposed a general pore anatomical parameter measurement method.Firstly,analyzing the structure of the Mask R-CNN,selecting a suitable feature extractor,and training the network model using self-made datasets and transfer learning methods.Secondly,using the trained model to segment the stomatal pores of the leaf microscopic image;each pore region is then extracted,whose boundary contour was fitted using the ellipse fitting algorithm based on the least squares method,the anatomical parameters of stomatal pore,including the major axis,minor axis,area and opening degree,etc.can be calculated;finally,adopting the evaluation indices such as precision,recall,and intersection of union(IoU)quantitatively evaluate the performance of the proposed method.This experiment verifies the accuracy of the method for the measurement of stomatal pore anatomical parameters.In order to improve the generalization ability of Mask R-CNN model,an improved Mask R-CNN stomatal pore segmentation method was studied.Combining edge detection technology,an edge extraction branch is added on the basis of the original Mask R-CNN network to enhance the network model's ability to extract pore edges.Among them,the edge extraction branch adopts the classical edge detection Sobel and Laplace operators based on the gaussian filter respectively to construct two improved Mask R-CNN stomatal pore segmentation models.The same training and testing data sets are used to evaluate the performance of the improved model from stomatal anatomical parameter measurement accuracy.The results show that edge information helps to improve the generalization ability of the network model.To meet the application requirements of rapid and accurate measurement of stomatal pore anatomical parameters,Yolact,a one-stage image segmentation network,was selected as the stomatal segmentation model in this paper.By using the feature map acquiring from fully convolutional neural network and mask coefficients to obtain the stomatal structure information and the complexity of the model is reduced.Although this method has a loss in the accuracy of stomatal pore measurement,its testing speed is faster than Mask R-CNN,and it can realize fast,accurate and high-throughput processing of stomatal pores.