The Reconstruction Method of SAR Image Ambiguous Area based on Deep Learning

Abstract

The ambiguities of Synthetic Aperture Radar (SAR) images have a serious impact on image quality and object detection accuracy. In recent years, Generative Adversarial Network (GAN) models based on Convolution Neural Networks (CNN) have been widely used in image generation and inpainting in the field of Computer Vision, making it possible to reconstruct SAR ambiguous regions using Deep Learning technology. Based on deep learning technology, the SAR image ambiguous area reconstruction method adopts the GAN model, which can achieve an ideal reconstruction of the marked area with SAR ambiguities. Compared with the SAR image ambiguous area reconstruction method based on traditional image processing, new method based on GAN has faster processing efficiency and better reconstruction result, can fully extract the image information, and learn the reconstruction from a large scale of area within complex background. To deal with the problem that the image reconstruction model based on deep learning has limited Receptive Field and difficult to reconstruct a large area of image, this paper introduces a conditional GAN structure based on the Fast Fourier Convolution layer (FFC layer) and studies the effect of its improvement to the feature extraction ability after the introduction of FFC. Finally, using a large-mask inpainting model based on FFC structure, a SAR image ambiguous area reconstruction method based on deep learning is purposed in this paper.