Research On Visual Recognition Method Of Grafting Seedling Image

Traditional artificial grafting not only has high labor intensity and low production efficiency,but also directly affects the efficiency and quality of grafting operation because of different human subjectivity and operation proficiency.With the development of mechanization and intelligence,mechanical grafting technology has been gradually used in agriculture,and the research on visual recognition has also been developed.The visual recognition method can accurately identify the target and obtain the obstacle information,which is the key to ensure the accurate identification and positioning of seedlings.This thesis focuses on image processing,and studies a visual recognition method based on image processing,which can realize the recognition of seedlings before grafting,so as to meet the requirements of agricultural operations.This thesis mainly studies the following aspects:(1)Image collection and preprocessing of seedlings.The collected seedling images mainly include time,shooting angle and height,and lighting conditions.The seedling image is preprocessed,and then the color component differences in different color spaces are analyzed according to the color differences the sample image.Finally,OTSU method and K-means method are used to segment the sample image to separate the target from the image background.(2)Extraction of key parameters of seedling samples.According to the requirements of agricultural operations,the two key parameters of cutting point and cutting angle are determined.Based on the analysis of the data of seedling stem diameter,true leaf,cotyledons,root position,horizontal intercept,etc.,the extraction method of key parameters was studied,the influencing factors of errors were analyzed,and the cutting point and cutting angle of seedlings were determined.(3)Analyze of convolution neural networks.By analyzing the principle,theory,structural characteristics and parameters of convolutional neural network,this thesis has a preliminary understanding of convolutional neural network.Then,the typical representatives of network models such as Fast R-CNN,Dense Net and YOLO algorithm model are analyzed and summarized.Finally,YOLO algorithm model is adopted as the main algorithm in this thesis and improvement measures in three aspects are proposed.(4)Realize seedling sample image recognition.In this thesis,the improved YOLO algorithm is used to classify and recognize sample images.Compared with traditional methods,the improved YOLO algorithm has been improved in recognition rate and accuracy to a certain extent by analyzing the algorithm complexity of the improved algorithm,Fast R-CNN and YOLO.