Study On Stoma Detection Method Of Living Plant Leaves Based On Faster R-CNN

Plant stomata is the channel of material and ability exchange between plant and environment,which plays an important role in regulating the carbon,water cycle and the mutual regulation between plant and environment.The accurate identification and analysis of stomata is the key to the calculation of stoma pattern and characteristic parameters.At present,the main method of collecting stoma image of plants is to tear the epidermis of leaves and make the specimen to be imaged under microscope.The statistical analysis of stoma quantity mostly adopts manual measurement or semi-automatic technology,which is difficult to achieve accurate,high-throughput and automatic processing.Therefore,based on the in-depth analysis of the characteristics of leaf stoma micro image,this paper constructs a stoma recognition network model based on Faster R-CNN target detection framework,optimizes the network parameters,and realizes the stoma detection and analysis of plant leaf micro image.In this paper,VHX-2000 digital microscope was used to collect more than 1000 stoma images of Poplar leaves at two magnification(500X,1000x).After preprocessing and target annotation software,three kinds of data sets including 500X stoma image,1000x stoma image and two kinds of ratio stoma image are made,and three kinds of stoma target detection models are trained respectively.Three models are used to cross detect 200 stoma images(100 images of 500X and 1000x respectively)containing two kinds of magnification.At the same time,the total number of stoma detected is automatically counted,and the accuracy and recall rate of stoma target detection are calculated.The accuracy of porosity detection is 99.92%,the highest recall rate is 99.32%(the same magnification data of 1000x model test),and the lowest is 89.59%(1000x data of 500X model test).According to the model with the highest recall rate and the stoma image parameters,the stoma density of the lower epidermis of Poplar leaves was 1.3/mm2.The experimental results show that the proposed model has the advantages of fast output and high accuracy.The stoma target detection model trained by Poplar stoma images was used to detect 100 stoma images of White Birch leaves,and the recall rate of the stoma target detection of White Birch was 95.60%.The results show that the proposed model has good generalization ability.