Estimation Of Leaf Shape Parameters Of Epipremnum Aureum Based On Transfer Learning

Epipremnum aureum is a perennial evergreen vine that is also known as the flower of life because of its tenacity and ability to come to life in water.Cultivating them indoors not only absorbs dust,purifies the air,and absorbs harmful substances,but also has a high ornamental value.Due to the many advantages of greenery,the public demand for it is increasing,so the continuous upgrading of greenery cultivation and management techniques and the formation of a complete set of standardized production techniques for greenery will achieve significant economic benefits in the practical application of production.Plant leaves are an important part of the external form of the plant,but also the main organ of the plant to carry out physiological functions.Accurate measurement of leaf length,width,leaf area and other shape parameters is important for understanding crop growth conditions and guiding crop breeding and cultivation.The current methods for measuring the shape parameters of plants or organs are mainly based on binocular stereo vision techniques or the combination of multi-view stereo vision and motion recovery structures,but these methods are timeconsuming and laborious in terms of camera calibration,iterative calculation,or multi-angle photography.In this paper,we propose a migration learning-based method for leaf shape parameter estimation,which provides a new way of thinking for plant or leaf shape parameter estimation.The main research contents and results of this paper are as follows.(1)Data acquisition and model library construction: The experiment uses the Kinect V2 camera to photograph the green lily plant from a single angle,and collects a total of 300 green lily leaf point cloud data and its corresponding numbered leaf shape parameter values,and carries out pre-processing operations such as filtering,segmentation and streamlining on the collected leaf point cloud.The simulated leaf model is constructed by inverting the parameter equations of the leaf shape parameters and calculating the leaf length,leaf width and leaf area shape parameters of the simulated leaf model.The simulated leaf model constructed by combining different parameters is discretized into point cloud data,and a point cloud model library containing 12763 simulated leaves is constructed in total.(2)Simulated blade shape parameter estimation based on multi-resolution coded point cloud network: In order to solve the problem that the maximum pooling operation in Point Net network causes the blade point cloud local features cannot be extracted well,the experiment proposes a multi-resolution coded network structure with iterative farthest point sampling algorithm to sample 64,128 and 256 points,respectively,for feature coding of point cloud data,so as to perform better capture of local features of the blade point cloud.Also compared with the single-layer perceptron of the original Point Net network structure,the experimentally proposed multilayer depth feature fusion perceptron performs better in the estimation of blade shape parameters.The R2 of the MRE-Point Net network is 0.9718,0.9788 and 0.9756 for leaf length,leaf width and leaf area,respectively,and the RMSE is 0.4813 cm,0.1894 cm and 2.7974 cm2,respectively.(3)Occluded leaf complementation based on multi-scale feature extraction module combined with point cloud pyramid decoder network: In order to solve the occlusion problem that exists in the green leaf photographed from a single angle by Kinect camera.A leaf shape complementation network based on multi-scale feature extraction module combined with point cloud pyramid decoder is proposed experimentally.The 11487 simulated leaves in the simulated leaf point cloud model library are used as the training set,the remaining 1276 simulated leaves are used as the validation set,and the 52 single complete point clouds of green lilies acquired by the Kinect camera and the 50 leaf point clouds and their corresponding numbered complete leaf point clouds acquired in the natural occlusion state are used as the test set.The 52 single complete point clouds were randomly mutilated at 20%,30% and 40%,and their complementary evaluation index MMD-CD values were 0.0069,0.0076 and 0.0081,and their complementary evaluation index MMD-EMD values were0.0082,0.0089 and 0.0093,respectively.The minimum,maximum and average values of MMD-CD were 0.0039,0.0087 and 0.0057,respectively.(4)Model migration-based estimation of green leaf shape parameters: In order to solve the problem of large workload of green leaf data collection and small training samples,a migration learning-based estimation method of green leaf shape parameters is proposed based on the training model for simulated leaf shape parameter estimation.A total of 300 point clouds and shape parameters of green leaves were collected,200 of which were used as the training set to fine-tune the pre-training model MRE-Point Net for model migration,and the remaining 100 were used as the test set to evaluate the ability of the fine-tuned model to estimate the shape parameters of green leaves.The results were compared and analyzed before and after the model migration,and the final mathematical statistics and linear regression analysis were performed between the estimated and true values of the shape parameters,resulting in R2 and RMSE of 0.9148 and 0.3903 cm for leaf length estimation,0.9225 and 0.2997 cm for leaf width estimation,0.9532 and 3.6454 cm2 for leaf area,respectively.