Abstract:  Point cloud data can accurately and intuitively reflect the spatial relationship between apples and branches. Because of their regularity of point cloud data, traditional convolution neural network is not suitable for point cloud data. Therefore, a semantic segmentation method of apple point cloud based on improved dynamic graph convolution is proposed. Based on Dynamic Graph Convolution Network (DGCNN), K-Nearest Neighbors (KNN) of different scales are used to construct the neighborhood relationship of each node. Adding neighbor node information in the edge convolution (EdgeConv) is to extract more abundant local features. A graph based attention module is designed to assign different weights to the K nearest neighbor points of the center point. Compared with using maximum pooling to aggregate features, this attention module can better aggregate the feature information of the neighborhood. Channel attention module is introduced to assign different weights to different features. The experimental results show that the network has a higher point cloud segmentation accuracy on the apple point cloud dataset, and the overall accuracy of OA and the average intersection ratio of MIOU reach 91.2% and 69.2%, respectively, OA and MIOU are 3.9% and 3.6% higher, compared with DGCNN.

Key words: apple classification, DGCNN, semantic segmentation, edge convolution, mixed attention mechanism

摘要： 点云数据可以准确、直观地反映苹果与树枝之间的空间关系，由于点云数据的不规则性，传统的卷积神经网络不适用于点云数据。因此，提出一种基于改进动态图卷积的苹果点云语义分割方法。基于动态图卷积DGCNN，采用不同尺度的K最近邻KNN构造各节点的邻域关系；在边缘卷积EdgeConv中加入邻居节点信息，提取更加丰富的局部特征；设计基于图的注意力模块，给中心点的K个最近邻居点分配不同的权重，相对于使用最大值池化对特征进行聚合操作，该注意力模块能更好地聚合邻接域特征信息；引入通道注意力模块，给不同特征分配不同的权重。试验表明，在苹果点云数据集上，该网络有较高的点云分割精度，整体精度OA和平均交并比MIOU分别达到91.2%和69.2%，相较于DGCNN，OA和MIOU分别提高3.9%、3.6%。

关键词: 苹果分类, DGCNN, 语义分割, 边缘卷积, 混合注意力机制