SAR target recognition based on model transfer and hinge loss with limited data
Qishan He; Zhao, Lingjun; Kuang, Gangyao; Liu, Li (2022-01-01)
He Q., Zhao L., Kuang G., Liu L. (2021) SAR Target Recognition Based on Model Transfer and Hinge Loss with Limited Data. In: Fang L., Chen Y., Zhai G., Wang J., Wang R., Dong W. (eds) Artificial Intelligence. CICAI 2021. Lecture Notes in Computer Science, vol 13069. Springer, Cham. https://doi.org/10.1007/978-3-030-93046-2_17
© Springer Nature Switzerland AG 2021. This is a post-peer-review, pre-copyedit version of an article published in Lecture Notes in Computer Science. The final authenticated version is available online at https://doi.org/10.1007/978-3-030-93046-2_17.
Convolutional neural networks have made great achievements in field of optical image classification during recent years. However, for Synthetic Aperture Radar automatic target recognition (SAR-ATR) tasks, the performance of deep learning networks is always degraded by the insufficient size of SAR images, which cause both severe over-fitting and low-capacity feature extraction model. On the other hand, models with high feature representation ability usually lose anti-overfitting capability to a certain extent, while enhancing the network’s robustness leads to degradation in feature extraction capability. To balance above both problems, a network with model transfer using the GAN-WP and non-greedy loss is introduced in this paper. Firstly, inspired by the Support Vector Machine’s mechanism, multi-hinge loss is used during training stage. Then, instead of directly training a deep neural network with the insufficient labeled SAR dataset, we pretrain the feature extraction network by an improved GAN, called Wasserstein GAN with gradient penalty and transfer the pre-trained layers to an all-convolutional network based on the fine-tune technique. Furthermore, experimental results on the MSTAR dataset illustrate the effectiveness of the proposed new method, which additional shows the classification accuracy can be improved more largely than other method in the case of sparse training dataset.
- Avoin saatavuus