Research Of Machine Visual Technology For Grape Yield Estimation

Precision agriculture is based on the Internet of Things.It controls the crops’ production process by deploying sensors and researching artificial intelligence technology,mainly including agricultural yield estimation,crop intelligent disease detection and growth process quality monitoring.Improves the fruit yield level at a lower cost.Among them,the fruit yield estimation is the most critical part of precision agriculture.At present,the work of grape visual yield estimation is mainly carried out from two aspects:detection-based grape cluster detection and regression-based grape berry counting.However,these two kinds of methods are usually relatively independent,and the calculation results are relatively one-sided with large yield estimation error.Therefore,a two-stage grape yield estimation framework based on cluster-berry fusion is proposed.The main contributions of this paper are as follows:(1)A set of two-stage non-destructive grape yield estimation framework based on computer vision is proposed.For the input grape image,the first stage is responsible for the locating and segmentation of grape clusters,while the second stage is responsible for the locating and counting grape berries.Two stages are correlated by the filter method based on Mask.Finally,in order to improve the computing efficiency,the image input is carried out by Patch.Summarize and analyze the existing work related to grape yield estimation and public datasets,and we transfer the deep-learning model of crowd counting based on traditional computer vision,regression and location to the grape field.We compare models effectiveness and efficiency at the cluster-level and berry-level respectively.(2)Focusing on serveral shortcomings in current grape yield estimation methods based on machine vision,a two-stage non-destructive yield estimation framework TSGYE is proposed,which fully integrates the characteristics of grape cluster and berry,achieving lower yield estimation variance.Otherwise,patch image is also considered for improving calculation efficiency.(3)In the second stage of TSGYE,regression-based grape berry counting models always receive low accuracy in grape density estimation,the author proposed a cascaded network based on GBCNet,the network was named as Joi Net,which can achieve more accurate berry counting results in the experiments.At the same time,a location-based grape counting method was proposed,which named LSC-CNN,and LSC-CNN improves the robustness of multi-scale branch fusion.According to the distance between the grape and its nearest neighbor,LSC-CNN generated the bounding box prediction for each grape instance,so as to better assist the display of grape counting result.(4)For the shortage of grape berry counting annotations,the author completed point-level annotations on the existing datasets WGISD and ICRA2015,and provided grape berry location informations.Both of these annotated datasets were published on Github websites,and were merged into the project by original author.(5)Deploy the two-stage grape yield estimation network on the mobile robot platform to demonstrate the efficiency and effectiveness of TSGYE.