| Summary: | Synthetic aperture radar (SAR) imaging has recently attracted considerable attention due to its variety of applications in both military and civilian aspects. However, a SAR image scheme can be affected by various elements that can lead to poor image reconstruction performance, especially for the target recognition mission; for instance, the complex environment, irregular sampling intervals, sample scarcity, imaging parameters, etc. The rapid development of deep learning currently makes it a great solution to deal with the aforementioned problems. In this paper, we propose a SAR image model based on conditional generative adversarial networks (cGAN), which combines image-to-image translation (pix2pix) and residual networks (ResNet) in order to diminish the noise and artifacts on SAR images, increase their signal-to-clutter-noise ratio (SNCR) of the images, and improve the short-range target recognition rate. Unlike conventional cGAN, we employ a ResNet-based discriminator (RbD) to effectively improve the SAR image denoising ability of the model. On the other hand, another similar discriminator is simultaneously trained to classify 14 familiar metallic object types with high accuracy and avoid the over-fitting problem. This discriminator is built by replicating the RbD one, and then we replace the last layer with the standard softmax function to classify multiple objects based on class probability outputs. The experiment results in this paper illustrate that the proposed scheme achieves higher image denoising performance and SNCR enhancement than the other conventional approaches. Besides, the target recognition rate of the proposed scheme outperforms the other common classification models.
|
|---|