Point cloud semantic segmentation technique is one of the effective means for point cloud processing and 3D scene understanding and analysis. For the problem that the local morphology in point cloud scenes varies, which makes it difficult for network model to recognize features, the method of neighborhood distribution relationship learning and mixed-scale fusion is proposed to enhance local perception. The idea of convolution operator is adopted, and then the correlation of high-dimensional features is learned based on the joint distribution relationship of all points in the neighborhood in the xyz directions to capture the overall local correlation. In addition, neighborhood fusion of a small range of underlying features and a large range of deep features effectively preserves the advantages of each part and assists the network in correcting similar or erroneous features. Finally the network architecture is established for the semantic segmentation task. Experiments were conducted on the datasets S3DIS and ScanNet, the results show that the method has improved in both overall and class-average accuracy evaluation metrics, proving its effectiveness.