Study On Maturity Discrimination Of "Red Globe" Grape Cluster Based On Convolutional Neural Network

“Red Globe” grape is a high-quality variety of table grape,and the output ranks first among productions in Chinese grape.“Red Globe” grape is a non-climacteric fruit,so it has no post-ripening period after harvest.Thus,it is most important for “Red Globe” grape to select appropriate harvest time.The grape clusters often were harvested too early or too late due to misjudgment of the maturity.Therefore,there is an urgent need to quickly and accurately judge the grape cluster maturity to increase the commodity rate of “Red Globe” grape.The image method was used in this study to determinate the maturity of “Red Globe” grape cluster in the natural environment.Firstly,“Red Globe” grape clusters were identified by different transfer learning models of convolutional neural networks.The background of the identified cluster images was segmented by the appropriate method.Then,the berries were detected by circular Hough transform(CHT)and berry maturity was determined by color characteristics.Finally,we developed an algorithm for determining the maturity of “Red Globe” grape accurately.The main results obtained in this study are as follows:(1)The transfer learning models were constructed using Fast R-CNN combined with different types of convolutional neural networks.The recognition model with VGG16 neural network obtained the best result.When the learning rate is adjusted to 0.0001 and the epoch is 20 times,the accuracy of model is up to99.07% with only 26 ms of the average detection time.(2)The K nearest-neighbor(KNN)algorithm is most suitable for background segmentation of grape cluster images.When the nearest number K is 5,KNN algorithm obtained the accuracy of 84.61% using Mahalanobis distance method.(3)The Log operator was selected to extract the edge of the first gradient image of the grape cluster.According to the round features of grape berry,the CHT method is used to detect grape berry with 0.15 of edge threshold and 0.94 of sensitivity It can detect more grape berries with 96.56% of accuracy at a higher speed.(4)The berries can be divided into four maturity levels according to the H value of the pixels from“Red Globe” grape image in the HSV space.Our developed algorithm can detect the grape cluster maturity with the overall accuracy of 91.14%.Also,the grape clusters can be classified into four maturity grades by our algorithm.The method proposed in this research can realize the maturity detection of the grape clusters in grapery.It can provide guidance for harvesting grape at the appropriate period in improving the commodity rate of “Red Globe” grape.On the basis of this study,it is also helpful for automatic picking of “Red Globe” grape.