SAR MARITIME OBJECT RECOGNITION BASED ON CONVOLUTIONAL NEURAL NETWORK
Yihang Zhi, Bing Sun, Yi Xu, Jingwen Li · IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium · 2022
Insufficient data of SAR target recognition task leads to low accuracy and poor generalization of model and the SAR imaging mechanism leads to the insignificant difference between ship targets, which make recognition difficult. To overcome the above problems, we propose a SAR maritime method using siamese networks for model pre-training. Siamese network produce sample pairs to ease training sample insufficiency, and output difference of sample pairs to help model learning heterogeneous difference. Then, transfer the pre-training parameters of feature exaction layer to an end-to-end model. Finally, the end-to-end convolutional neural network is obtained by fine-tuning the parameters with supervised information. Experimental results show that the SAR maritime target recognition method based on siamese network training can effectively improve the recognition accuracy under the training condition of a small number of samples.