Research On Grading And Appearance Feature Detection Method Of Grafting Melon Rootstock Seedlings

In order to achieve the goal of classifying and extracting appearance parameters of seedlings for grafting of cucurbit crops,this paper applies machine vision and deep learning technologies to grafting technology,and takes the "Jingxin Zhin 4" pumpkin rootstock seedling as the research object,detecting and classifying artificial greenhouse-cultivated pumpkin rootstock seedlings.A seedling grading recognition network model was established,and a seedling grading recognition inference model was trained using a self-constructed pumpkin hole tray seedling dataset.Furthermore,appearance features were extracted from pumpkin rootstock seedlings that met the grafting conditions,and a pumpkin rootstock seedling appearance feature extraction method based on deep learning was proposed.Parameters such as the growth point position,leaf expansion direction angle,and petiole expansion angle of pumpkin rootstock seedlings under different perspectives were extracted and compared with manually measured values,verifying the feasibility of the proposed method.The main research content and conclusions are as follows.:(1)Conduct seedling experiments to cultivate the required pumpkin rootstock seedlings,measure some of the external parameters of the pumpkin rootstock seedlings,and according to the research purposes,collect pumpkin rootstock seedling image data from the tray and single plant pumpkin rootstock seedling image data from different angles.(2)The grading standard for pumpkin rootstock seedlings in the plug tray was proposed,and the pumpkin rootstock seedling dataset was created in PASCAL VOC format.The self-made dataset was used to train an improved YOLOv5 s network,and the pumpkin rootstock seedling classification recognition model was obtained and compared with other models.Experimental results show that the improved YOLOv5 s network proposed in this paper performs the best in the pumpkin rootstock seedling classification recognition task,with mAP＿0.5 reaching 99.1% and mAP＿0.5:0.95 reaching 80.6%,meeting the experimental requirements.(3)The segmentation dataset of single pumpkin rootstock seedlings under different viewing angles was created,and the self-made dataset was trained using the Mask-RCNN network to obtain the pumpkin rootstock seedling segmentation model.After experimental evaluation,the trained pumpkin rootstock seedling segmentation model achieved 100% accuracy in identifying true leaves and cotyledons based on the overhead viewing angle,and achieved 82.9% accuracy in segmenting the growing point and an average segmentation accuracy of 94.3%.For seedlings viewed from a side angle,the accuracy of identifying true leaves,cotyledons,growing points,and stems were 97.9%,98.7%,95.9%,and 100%,respectively,with an average precision of each category of 98.1%.The experimental results demonstrate that the trained seedling segmentation model meets the experimental requirements and can successfully complete the critical part segmentation task of seedlings.(4)The position of the pumpkin rootstock seedling’s growth point and the angle of the cotyledons’ unfolding direction were extracted from a top-down perspective,and the position of the growth point and the angle of the leaf stem’s unfolding were extracted from a side view.These values were compared with manually measured values,and the experimental results showed that the method proposed in this paper had a maximum error of 2 pixels and an average error of 0.6 pixels for the recognition of the growth point position of the pumpkin rootstock seedling from a top-down perspective.For the recognition of the angle of the cotyledon’s unfolding direction,the maximum error was 5°,and the average error was 1.7°.For the recognition of the growth point position of the pumpkin rootstock seedling from a side view,the maximum error was 4 pixels,and the average error was 1.2 pixels.For the recognition of the angle of the leaf stem’s unfolding,the maximum error was 7°,and the average error was 2.9°.These values met the accuracy requirements for grafting.Therefore,the proposed method of detecting the appearance characteristics of seedlings has certain feasibility and can provide some scientific research value and practical significance for the development of automatic grafting technology.