SAR Image Ship Object Generation and Classification With Improved Residual Conditional Generative Adversarial Network
Lu Li, Chao Wang, Hong Zhang, Bo Zhang
Abstract
Image generation by the conditional generative adversarial network (CGAN) can provide rich samples for supervised classification tasks. However, the unstable gradient updating problem makes it difficult to be used for image generation, especially for the high-resolution SAR image. In this letter, an improved residual condition generation network is proposed to enhance the quality of generated images and classification accuracy for hard negative examples. First, a residual convolutional-based block is built to clear the detail texture of different types of targets. Then, the gradient penalty and Wasserstein loss are used as discriminators to improve the similarity between real samples and the intra diversity of a generated images. We test the proposed method on generation and classification of three kinds of commercial ships using C-band 3-m Gaofen-3 (GF-3) SAR images. Experimental results show that our method can generate high-quality ship samples and achieve good classification accuracy.