Phenotypic Characteristics Extraction Of Grape Fruit Based On Deep Learning

Accurate and timely irrigation according to the water deficit status of grapes can not only reduce irrigation water consumption,reduce production costs,but also increase yield and improve quality.Fruit is an important reproductive and vegetative organ of grapes.Studies have shown that phenotypic characteristics such as fruit size and cluster color can accurately reflect the growth and water deficit status of grapes.Accurately extracting phenotypic characteristics of grapes can sense the water deficit status of fruit trees.However,due to the phenomenon of overlapping between the grapes and the interference of non-target objects such as poor lighting conditions and branches and leaves in complex agricultural environments,it is difficult to effectively extract the phenotypic characteristics of the fruits based on traditional machine vision methods.In this paper,focusing on the measurement of the geometric feature parameters of the grape fruit and the extraction of the color feature of the grape cluster,based on the deep learning method,the key algorithms such as fruit contour extraction,contour fitting and grape cluster semantic segmentation are studied.The main contents are as follows:Aiming at the problem of difficulty in obtaining phenotypic characteristics caused by the overlapping growth of grapes,the research proposed an algorithm for extracting the contour of the fruit based on the improved HED network.The algorithm is based on the convolutional neural network,which realizes the automatic extraction of the edge feature of the fruit and predicts the contour.Compared with traditional edge detection algorithms,the improved algorithm accuracy and robustness under complex environments have been greatly improved.Based on the extraction of the contour feature of the fruit,in order to solve the obstacles caused by the lack of the contour of the fruit to the automatic measurement of the geometric parameters of the fruit,a contour fitting algorithm of the grape fruit based on the candidate region is proposed,and the fruit is extracted by the YOLOv3 network,an iterative least-squares ellipse fitting algorithm is proposed to fit the contour in the candidate area,accurately and effectively restore the original contour of the fruit,so as to achieve accurate measurement of the geometric parameters of the fruit.Aiming at the problem of grape cluster region extraction and color feature analysis in a complex agricultural environment,the study proposed a multi-source image semantic segmentation algorithm based on deep learning.The dual-stream network structure was used to extract the RGB image and depth map features separately,and multi-scale feature fusion was carried out,exploring the correlation between different modalities,and has realized the effective extraction of regional characteristics of grape cluster.On this basis,an automated analysis and measurement platform for grape fruit phenotypic characteristics was designed and developed,and related algorithms were verified.In this paper,the grape fruit image segmentation algorithm developed based on the deep learning framework is helpful to overcome the difficulty of extracting the phenotypic characteristics of the overlapping fruit in the complex environment,and realizes the precise stress irrigation of the grape cultivation,thereby reducing the cost of grape cultivation.It is of great significance to improve the yield and quality of grapes.