Algorithm For Monitoring Growth Of Cabbage Seedlings Based On Fusion Of Spatial And Temporal Features

Cabbage is a kind of cruciferous plant,which is widely grown in temperate and subtropical areas of our country.At present,our country mainly uses the method of growing cabbage seedlings by plug trays in greenhouses.Most intensive nursery factories still use manual methods to identify the growth of cabbage in the seedling stage,lacking efficient and convenient automatic growth monitoring methods,and it is difficult to meet the development needs of intelligent agriculture and precision agriculture.In view of the above problems,this thesis took the collected image data of plug cabbage seedlings as the research object,and studied the algorithm for monitoring growth of cabbage seedlings based on fusion of spatial and temporal features,aiming to improve growth analysis performance throughout the entire seedling cycle of cabbage.This study promoted the development of intelligent cabbage seedlings management technology of greenhouse.The main research content of this thesis is summarized as follows:(1)Segmentation algorithm of cabbage seedlings based on YOLACT-RFX model.This algorithm was improved on the basis of the YOLACT instance segmentation algorithm,which is used to study the segmentation algorithm suitable for cabbage seedlings in different growth stages.This study can provide technical support for the subsequent phenotype analysis of cabbage during the whole seedling cycle.In order to improve the segmentation performance and seedling stage recognition ability of densely planted cabbage seedlings in plug trays,the recursive feature pyramid structure was introduced to strengthen the extraction of feature information at the edge of the leaves of adjacent hole positions.Aiming at the characteristics of rapid size change during the growth of seedlings,the atrous spatial pyramid pooling structure was introduced to enhance the recognition ability of cabbage seedlings at different seedling stages.Finally,the Res Ne Xt backbone network was used to speed up model convergence while reducing network parameters.The experimental results showed that when the intersection ratio threshold was 0.5,the improved YOLACT-RFX algorithm achieved an average precision of 84.4% and an average recall rate of 92.7%,which were respectively improved by 3.6% and 3.9% compared with the original YOLACT algorithm.Compared with the mainstream MASK-RCNN,SOLO and Query Inst segmentation algorithms,the average accuracy of YOLACT-RFX algorithm has increased by 2.5%-22%,and it has better seedling stage recognition ability and segmentation performance.(2)Phenotype analysis method of cabbage seedlings based on binocular feature matching.This method was based on the estimation method of growth indicators such as plant height,plant width and leaf area based on the images of binocular cabbage seedlings plug trays.The growth parameters were calculated by combining YOLACT-RFX model and binocular image feature matching algorithm.By introducing the Lo FTR feature matching algorithm to identify the feature points of the cabbage seedlings in the corresponding holes,and using the instance segmentation mask area to constrain the feature points,combined with the principle of binocular distance measurement to realize the plant height calculation.The plant width estimation was realized by calculating the mean value of the maximum side length of the object detection boxes in the binocular image.Finally,the seedling leaf area was estimated by counting the number of pixels in the instance segmentation mask.The experimental results showed that under the condition of spaced planting,the correlation coefficients between the estimated plant height and plant width and the manual measurement were 0.91 and 0.92,respectively;under dense planting conditions,the corresponding correlation coefficients were0.86 and 0.85;In the leaf area estimation experiment at different seedling stages,compared with the number of mask pixels manually marked,the accuracy rate of interval planting was above 89.4%,and the accuracy rate of dense planting was above 81.6%.This phenotype analysis method can realize the estimation of growth parameters of cabbage at different seedling stages,and has certain practical value.(3)Based on the Python development language,the Py Side2 image interface development component and the Py Torch deep learning framework,a cabbage seedlings growth monitoring system was designed and implemented.The functions covered model training,seedlings emergence detection,seedling identification and growth parameter estimation.This system integrated the algorithm model involved in this thesis,with friendly interactive logic design and simple and easy-to-understand interface,which can meet the needs of growth monitoring in the process of growing cabbage seedlings in greenhouses.