Research On Nondestructive Measurement Algorithm For The Watermelon Plug Seedlings Phenotype Based On RGB-D Camera

The seedling phenotype measurement mainly relies on manpower,which is inefficient and prone to seedling damage.Also,the measurement limits the research that is based on seedling phenotype data.It is urgent to replace manual work with Highthroughput non-destructive measurement technology.The paper,taking the whole-tray of watermelon seedlings as an example,proposes a non-destructive measurement method for the phenotype of the whole-tray of watermelon seedlings from 1 true-leaf stage to 3true-leaf stage,and realizes the real-time measurement of phenotypic data in different growth periods.And the measurement accuracy meets the use requirements.It provides a useful example for phenotype detection-related research of whole-tray seedlings,which is expected to become a measurement tool for seedling related research and has a good application value.(1)The whole-tray seedling image collection work was carried out by the Azure Kinect camera.According to the collection requirements of the whole-tray seedling image,the image acquisition device was produced,and two Azure Kinects were installed on its top and side positions which respectively were aimed to shoot Sequences of canopy and stalk image of whole-tray of seedlings.The automatic/semi-automatic image acquisition software used in conjunction with the image acquisition device has been developed,which can display the working status of the camera and realize real-time acquisition of color image,depth image,color-aligned depth image(RGB-D)and depth-aligned color image(D-RGB)data,significantly improving the acquisition efficiency.A deep learning dataset,which used for the collected original images,including the recognition of growing point image,the annotation of leaf image and stalk image segmentation,was created.Also,the annotated images are expanded to provide data support for later deep learning models.(2)Algorithms for seedling growth point recognition,leaf and stem image segmentation was proposed based on deep learning.For the image recognition of the growth point of the whole-tray seedling,two target detection algorithms,YOLOX and Faster RCNN,were tested.Through comparative analysis of experiments,YOLOX had better performance for growth points,and the recognition accuracy was higher than96.8%.For the recognition and segmentation of the whole-tray seedling leaf image,two segmentation algorithms,Mask RCNN and YOLACT,were tested.After comparative analysis in the experiment,Mask RCNN had better performance,and the accuracy of leaf recognition,classification and segmentation was higher than 91.7%.For the image of the whole-tray seedling stalk,the Mask RCNN algorithm has achieved a good segmentation effect,and the segmentation accuracy is higher than 095%.The above seedling organ image recognition and segmentation algorithms meet the requirements of phenotypic measurement and provide accurate image segmentation effects for subsequent phenotypic measurement.(3)The calculation method of phenotypic parameters such as seedling number,plant height,real leaf number,cotyledon number,leaf area,stem diameter and hypocotyl was designed.According to the specification characteristics of the plug tray,combined with the image recognition and location information of the growth point,the growth point index method was established to determine the number of seedlings and the actual growth position of the seedlings in the plug tray.According to the definition of plant height,the height difference between the growth point and the plug tray substrate was calculated in the depth image.The Mask RCNN leaf image segmentation algorithm can output the category,quantity and segmentation results of leaves,and then use the Cycle-Consistent Adversarial Networks and image processing algorithm to restore the color and depth information of the leaf occlusion part,and realize the leaf area measurement in the leaf point cloud.The Mask RCNN stalk image segmentation algorithm can obtain the mask map of the stalk and extract its skeleton.Using the stem mask image and the skeleton pixel coordinates,the depth information in the corresponding depth map can be obtained,thus achieving non-destructive measurement of hypocotyl length and stem thickness.(4)To complete the performance evaluation of the phenotype detection algorithm for the whole-tray watermelon seedlings in different growth stages.In the measurement of canopy phenotype(leaf number,leaf area,growth point and plant height)and stalk phenotype(stem diameter and hypocotyl length)3 plates of seedlings were selected for the algorithm performance test.Then the algorithm measurement results were compared with manual measurement results.The experimental results show that the average measurement accuracy rate of plant height phenotype is 95.34%,the average measurement accuracy rate of true leaf number is 97.94%,the average measurement accuracy rate of cotyledon number is 95.31%,and the average measurement accuracy rate of leaf area is 91.70%.The average measurement accuracy rate of hypocotyl length is92.59%,and the average measurement accuracy rate of stem diameter is 91.79%.The measurement results demonstrate that the phenotypic non-destructive testing method proposed in this paper shows better measurement accuracy and meets the requirements of phenotypic testing.(5)A non-destructive measurement software for the whole-tray seedling phenotype was developed by using Py Qt.The software interface was optimized according to user habits.The deep learning network algorithms and phenotype calculation methods were integrated into the phenotype measurement software,at the same time,a standardized operation procedure was developed to simplify the tedious phenotype measurement process.