基于多尺度生成对抗网络的SAR图像样本增广

Data Augmentation in SAR Images Based on Multi-Scale Generative Adversarial Networks

针对军用无人飞行器对海上重要舰船合成孔径雷达图像获取困难的问题,提出了一种从单一图像学习图像内部分布的无条件图像生成网络。该网络采用金字塔式多尺度生成对抗网络(GAN)思想,在每一层金字塔中都有一个GAN负责该尺度下图像块的生成和判别,且每个GAN具有相似的结构。生成器前端采用不同大小卷积核连接的Inception模块获取不同尺度下的图像特征,为了充分利用这些特征,加入了残差密集模块;判别器采用马尔科夫判别器的思想,捕捉不同尺度下的图像分布。将所有生成的图像制成数据集用于训练不同的目标检测算法,结果表明,训练后模型的平均精度得到了一定的提升,验证了该网络模型的有效性。

To solve the problem that it is difficult for military unmanned aerial vehicles to acquire synthetic aperture radar images of important ships at sea,this paper introduces an unconditional image generation network which can learn the internal distribution of images from a single image.The network adopts the idea of a pyramid of multi-scale generative adversarial networks(GAN).In each layer of pyramid,there is a GAN responsible for the generation and discrimination of image blocks at this scale,and each GAN has a similar structure.The head of generator contains Inception modules connected with different sizes of convolution kernels to obtain image features at different scales.In order to make full use of these features,a residual dense block is added.The discriminator uses the idea of Markov discriminator to capture images distribution at different scales.All the generated images are made into data sets for training different target detection algorithms,the results show that the average accuracy of the model is improved to a certain extent,which verifies the effectiveness of the network model.