Convolutional neural network has made remarkable achievements in classification of idealized point cloud, however, non-idealized point cloud classification is still a challenging task. In this paper, DNDFN, namely, Dual-Neighborhood Deep Fusion Network, is proposed to deal with this problem. DNDFN has two key points. One is combination of local neighborhood and global neigh-borhood. nearest neighbor (kNN) or ball query can capture the local neighborhood but ignores long-distance dependencies. A trainable neighborhood learning meth-od called TN-Learning is proposed, which can capture the global neighborhood. TN-Learning is combined with them to obtain richer neighborhood information. The other is information transfer convolution (IT-Conv) which can learn the structural information between two points and transfer features through it. Extensive exper-iments on idealized and non-idealized benchmarks across four tasks verify DNDFN achieves the state of the arts.