Research On Picking Point Location Of Field Litchi Based On Convolutional Neural Network And Density Clustering

As a special fruit of Lingnan,litchi is known as the "Savoring hundreds of litchi is a day,I want to be a man in Five Ridges forever ",while the production of litchi is extremely time-consuming and laborious,especially the litchi picking,which greatly consumes labor.As the rural population shifts to the city,a large amount of labor flows into the city,resulting in a serious labor shortage in the crop industry,while the ripening period of litchi is very short.In addition,the hot and rainy areas in the south China will lose weight if not harvested in time,causing serious economic losses.Litchi picking robots can effectively solve problems such as labor shortage and large-scale planting,which is of great significance for reducing the production cost of litchi and alleviating the lack of productivity caused by the loss of agricultural population.Computer vision is an important part of picking robots,providing powerful technical support for fruit detection and positioning.With the development of deep learning and convolutional neural networks,deep learning has highlighted many advantages such as robustness and high precision in the field of computer vision.This paper uses the deep learning and clustering algorithm to study the location of picking points as follows:(1)Aiming at the complicated background conditions in the wild environment and the serious overlapping of the litchi fruits themselves,an improved target detection algorithm based on the YOLOV3 model was proposed for the detection of litchi fruits.Based on the idea of residual network and dense convolution convolution,a feature extraction network that combines dense convolutional blocks and residual blocks is proposed.Experiments show that the improved algorithm achieves better results than the classic YOLOv3 algorithm Tiny feature extraction network and Dark Net53 feature extraction network.The m AP of litchi fruit is 97.07%,which is higher than the YOLOv3＿Tiny’s 94.48% m AP,which is also higher than the classic YOLOv3.95.18% m AP.The time efficiency can reach 58 FPS,which is higher than the 29 FPS of the classic YOLOv3.(2)Aiming at the problem of difficult positioning of branches in the wild environment,this paper proposes a semantic segmentation of litchi branches using deep learning image segmentation algorithm,and uses Deep Lab V3+ framework and Xception＿65 feature extraction network to realize semantic segmentation of litchi branches.The model greatly reduces the operation parameters through the coding and decoding structure.By using the deep semantics and the shallow semantics to merge with the weights,the model has a greater degree of detail description.The model uses a porous whole convolution pool to reduce the precision loss caused by pooling without increasing parameters.Experiments show that the MIPo U obtained by Deep Lab V3+＿Xception＿65 model segmentation is 0.765,which is much larger than the segmentation accuracy of other comparison models and other feature extraction networks under the same architecture.Compared with the deep residual network and Mobile Net network under the same architecture,the experimental results are obtained.(3)Based on the characteristics of the growth of litchi clusters in the complex environment of the field,the density-based OPTICS clustering algorithm is used to cluster the distribution of fruits in the image for the litchi vision image with less fruit branches.Different clusters are used to better find the picking position of the whole cluster.For the near-field image,the single-fruit and multi-fruit cases are discussed separately,and the corresponding picking point localization algorithm is proposed.Finally,the paper summarizes the flow of the litchi picking process,and uses the clustering algorithm to conduct sampling experiments on the existing sample data.The experiment shows that in the extracted 100 sample images,the maximum circumscribed rectangle is obtained by combining semantic segmentation and target detection.The algorithm for obtaining the picking point can accurately locate 73 picking points.In the sample image of the foreground sampling,the clustering accuracy rate reaches 93.3%,and the experiment obtains better results.