| Seedling stage is a critical period for crop growth and development.The phenotypic data of seedling plant height and stem diameter are important basis for reflecting seedling growth and development.Non-destructive measurement of plant height and stem diameter of seedlings is of great significance for pest control of seedlings and breeding and screening of excellent crop varieties.The manual measurement method has a large workload,low efficiency,large error and is easy to cause irreversible damage to seedlings during the measurement process.The existing measurement methods have many problems,such as many calibration constraints,poor real-time computing and poor measurement effect of small targets.Based on deep learning technology,this study carried out a non-destructive measurement method of seedling height and stem diameter based on monocular depth estimation,in order to improve the phenotypic monitoring ability during seedling cultivation.The height and stem diameter data(2808 groups)of tomato,pepper and cabbage seedlings(216 seedlings)under different shooting distances,different light intensities,multiple seedling targets and different tilt angles were tested to verify the effect of the model.The main research contents and conclusions are as follows:(1)A monocular depth estimation method based on Convolutional neural network(CNN)was proposed.The encoder was replaced with Res Net,Dense Net and Res Ne Xt to explore the depth generation characteristics and laws of monocular images.Based on this method,a nondestructive measurement experiment of seedling height was carried out to verify the feasibility of monocular depth estimation in seedling height measurement.The test results on the NYU Depth V2 depth dataset show that the Average relative error(Rel),Root mean square error(RMSE),Average logarithmic error(LG)and accuracy(δ<1.25)of the monocular depth estimation model based on Res Ne Xt are 0.175,0.165,0.060 and 0.798,respectively.Compared with Res Net and Dense Net,Res Ne Xt was reduced by 72.9 % and 78.6 % respectively on RMSE,and 12% and 3.3% respectively on Rel.The accuracy(δ<1.25)increased by 32.1% and 21.6%respectively.The depth estimation model was used to validate the seedling height measurement,and the Mean absolute error(MAE)of tomato,pepper,and cabbage within 105 cm were 0.569,0.616,and 0.326 cm,respectively;RMSE＿D is 0.829,0.672,and 0.389 cm,respectively;The average ratio R is 1.005,1.001,and 1.012,respectively,indicating that the method is feasible and universal.When the number of plants in a single image was less than 5,the average values of R,MAE and RMSE＿D are 1.001,0.652 cm and 0.829 cm,respectively.When the sensitivity is less than 160,the MAE of plant height is 0.81 cm.Therefore,for multi-target and different light environments,the height measurement of the plant can be completed,which improves the practicability of the method.(2)Aiming at the global feature generation characteristics of existing monocular depth estimation tasks and the lack of local details in seedlings,a monocular depth estimation network architecture integrating Transformer and multi-scale local plane guidance(LPG)was proposed.The global depth feature information was obtained through the feature expression module Trans Block based on the Transformer mechanism and the adaptive channel attention module TGCA.Through the local information generation module multi-scale LPG and multi-scale feature fusion module FFBlock,the generation of local depth details and the adaptive fusion of different scale depth features were completed.On the NYU Depth V2 dataset,this model has achieved optimal performance on multiple evaluation indicators.Rel and RMSE decreased to0.095 and 0.346,respectively,and the accuracy(δ<1.25)increased to 0.918.On the KITTI dataset,this research model has achieved the most advanced results.Among them,Rel,RMSE and RMSE log decreased to 0.056,2.299 and 0.086,respectively,and the accuracy(δ<1.25)increased to 0.967.The experimental results showed that the depth estimation model in this study can well completed the depth estimation task of seedlings,and provided effective depth information for phenotypic calculation of seedling height and stem diameter.(3)In order to verify the actual effect of this method on the measurement of seedling height and stem diameter,this study set up comparative experiments with different shooting distances,different light intensities,multiple seedling targets and different tilt angles for tomato,cabbage and pepper seedlings with large phenotypic differences.The results showed that when the shooting distance was within 105 cm,the average MAE and RMSE＿D of seedling height were0.519 and 0.63 cm,respectively,and the average MAE and RMSE＿D of stem diameter were0.039 and 0.047 cm,respectively.Among all sensitivities,the average MAE of seedling height and stem diameter were 0.332 and 0.022 cm,respectively,and the average RMSE＿D were 0.401 and 0.026 cm,respectively.For multiple targets,the average MAE and RMSE＿D of seedling height were 0.501 and 0.602 cm,respectively,and the average MAE and RMSE＿D of stem diameter were 0.043 and 0.054 cm,respectively.When the tilt angle was within 15°,the average MAE and RMSE＿D of stem diameter were 0.017 and 0.024 cm,respectively.Therefore,the accurate measurement of seedling height and stem diameter was achieved under the above different conditions in this study,indicating the high practicality of the proposed method.(4)Utilizing the PyQT5 graphical interface design tool and Python programming language,combined with the proposed monocular depth estimation method that combined Transformer and multi-scale LPG,a demonstration software for non-destructive measurement system of seedling height and stem diameter based on monocular depth estimation is designed and developed.After experimental verification,the software can complete the automatic and rapid measurement of seedling height and stem diameter and the autonomous measurement of different seedling phenotypic data. |
