| abstract |
The invention discloses a transfer learning-based small sample recognition method for spatial objects with high recognition accuracy, which overcomes the problems of cumbersome manual feature extraction and feature engineering for spatial object recognition in the prior art. The invention contains the following steps, step 1, establishing an auxiliary sample space target data set; step 2, constructing an end-to-end deep nearest neighbor network; step 3, sending the auxiliary data set into a deep nearest neighbor network for training; step 4, constructing a space Target data set; step 5, send the target data set to the deep nearest neighbor network for identification. This technology uses two loss joint training, aiming at the recognition of fine-grained domains for spatial target recognition, small inter-class differences and large intra-class variance, and the introduction of intra-class compact constraints makes the same class of samples in the feature space as much as possible are similar, so that the present invention can still obtain good recognition results under the condition that the intra-class variance of the spatial object image is relatively large. |