Research On Recognition Method Of Tea Bud Based On Computer Vision

Tea planting has a long history in China,tea has become one of the main crops in China,and the planting area is expanding.With the continuous development of tea market economy,the proportion of profits brought by tea planting in farmers’ income is also increasing.The current way of picking tea is divided into two kinds,one kind is with the method of manual labor to harvest,because of manual picking has more complete tea buds,so it usually has higher economic benefits.However,this kind of picking method cannot meet the demand of tea production in the tea market because of low picking efficiency and insufficient labor force,therefore,another mechanical picking method has been developed to replace manual picking.The efficiency of mechanical picking is many times higher than that of manual picking,which can meet the large demand for tea production in the tea market,but most of the current picking machines in the market use one cut method to harvest the tea on the top of tea tree,the integrity of the tea shoots obtained by this method is not high,which will damage the quality of tea and reduce the economic benefits of tea.Therefore,in order to promote the automatic development of the tea industry and ensure the quality of tea,this thesis uses computer vision technology to identify the tea picking points based on the tea images in the natural background,which provides technical support for the mechanized tea picking in the future.In this thesis,the problem of tea picking under natural conditions is studied as well as proposed a method which based on Res Net and U-Net to segment tea image under the natural background,and achieved good segmentation results.Aim at detecting the picking points of single bud,one bud and one leaf,one bud and two leaves,SSD network is used to detect the picking points.In order to detect the picking points of tea buds more accurately,SSD model based on high-level and low-level feature fusion is proposed,which improves the detection accuracy of traditional SSD model and can detect the picking points of tea buds more accurately.The main research and results of this thesis are as follows:(1)Aiming at the tea image in outdoor environment and natural light,a method of tea bud image segmentation based on Res Net and U-Net is proposed to segment the tea bud in natural background.By constructing the binary image of tea bud and feeding it into the tea bud image segmentation model for training,the model can learn the rich semantic features of tea bud.The experimental results show that the proposed method can segment the tea bud image well in the natural background,and ensure the complete contour of the tea bud after segmentation.(2)Aiming at the problem of picking point detection of single bud,one leaf and one bud as well as one bud and two leaves,an improved SSD object detection model is proposed.By designing SSD model feature fusion method,the high-level feature map and low-level feature map of SSD model feature pyramid are fused,and the detection accuracy of each leaf picking point is improved.