In this paper, a point cloud semantic segmentation network combining multi-scale supervision and SCF-Net is proposed to address the problems of low segmentation accuracy of point cloud in complex scene, the lack of direct supervision in neural network hidden units, and the difficulty in extracting specific point cloud features. A category information generation module is first constructed to record the receptive field categories of hidden unit in the encoder, which is used for the supervised learning of auxiliary classifiers in the decoder. Secondly, the point cloud category prediction task in the decoding stage is decomposed into a series of point cloud receptive field category prediction tasks. By adding auxiliary classifiers to each layer of the decoder, the point cloud receptive field category of the current stage is predicted and the category information generated in the coding stage is used as the label to supervise network learning. The model infers point cloud receptive field categories from coarse to fine, and finally predicts point cloud semantic labels. The experimental results show that the method can effectively extract key information of point cloud and improve the accuracy of semantic segmentation.