Tea tender shoot joint segmentation method based on graph convolution

Abstract

The invention provides a tea tender shoot joint segmentation method based on graph convolution, and the method comprises the steps: carrying out the convolution operation of an input point through a multi-layer down-sampling graph, and obtaining instance feature information and semantic feature information through a linear interpolation up-sampling module and a gating propagation module; and performing feature fusion on the semantic feature information and the instance feature information in a cascade attention fusion module to obtain instance information containing semantic features and semantic information containing instance features. According to the method, the parameter quantity of the model is reduced while the point cloud edge feature information is improved, and the segmentation precision is improved.

Description

technical field [0001] The invention belongs to the technical field of point cloud segmentation, and in particular relates to a joint segmentation method of young tea shoots based on graph convolution. Background technique [0002] Due to the disorder and unstructured characteristics of point cloud data, the convolution operation of traditional two-dimensional images cannot be directly used in point cloud data processing, and the effect of class convolution improved by imitating two-dimensional convolution is far inferior to that of convolution. The effect of product neural network in the field of two-dimensional image processing. Although after the pointnet series of algorithms, many scholars have proposed point cloud convolution operations, such as pointcnn, pointconv and other algorithms; and recently discovered graph convolutional networks, such as DGCNN, etc., but there are still the following defects: ① When performing KNN Or when Radius NN searches, due to the differ...

AI Technical Summary

Problems solved by technology

[0002] Due to the disorder and unstructured characteristics of point cloud data, the convolution operation of traditional two-dimensional images cannot be directly used in point cloud data processing, and the effect of class convolution improved by imitating two-dimensional convolution is far inferior to that of convolution. The effect of product neural network in the field of two-dimensional image processing

Although after the pointnet series of algorithms, many scholars have proposed point cloud convolution operations, such as pointcnn, pointconv and other algorithms; and recently discovered graph convolutional networks, such as DGCNN, etc., but there are still the following defects: ① When performing KNN Or when Radius NN searches, due to the difference in the order of the search points, the disorder of the point cloud is caused. Most methods use the form of mlp, class convolution and maxpooling to extract features, and the extracted point features cannot obtain the local geometry of the point cloud. The interaction between points, and the point cloud edge information obtained by these feature extraction operations is relatively poor, which greatly affects the final segmentation accuracy; while in the segmentation of tea shoots, the edge information features affect the final segmentation effect. It has a great impact; ②At present, the downsampling operation is not introduced in the graph convolutional network, which will lead to too many parameters of the network model, which is difficult to be used for the deployment of embedded devices for real-time tea shoot segmentation tasks; ③Large Most networks combine semantic segmentation and instance segmentation tasks in series. The dependence between the two is too strong, and the training time of the network model of this branch line is relatively long, and the network convergence speed is slow.

Images