基于GAN的遥感图像协同数据增强方法

Collaborative Data Enhancement Method of Remote Sensing Images Based on GAN

针对基于深度学习的分类模型在少样本训练时所遭受的梯度消失、过拟合问题,结合DCGAN和SRGAN特性,提出一种抑制过拟合、提升图像生成质量的DS-GAN协同数据增强算法。通过改进DCGAN生成新的图像,使用改进SRGAN对其进行超分辨率重构,二者协同得到新的超分辨率图像。首先,提出一种软标签函数,代替DCGAN原始固定标签;其次,引入空洞卷积残差块作为SRGAN判别器主结构,同时加入CBAM注意力机制实现权重的再分配;最后,在SRGAN判别器中引入自适应平均池化,降低网络参数量。实验结果表明,使用标准数据集AID和RSOD,经MobileNet V2分类网络进行测验,DS-GAN数据增强方法相较于常规增强和DCGAN增强方法有明显提高。在AID数据集上,准确率分别提升8.01%、9.49%。在RSOD数据集上,准确率分别提升4.76%、1.4%。

Aiming at the problems of gradient disappearance and overfitting suffered by classification models based on deep learning when training with few samples,combining DCGAN and SRGAN characteristics,a DS-GAN collaborative data enhancement algorithm is proposed to suppress overfitting and improve image generation quality.By improving DCGAN to generate a new image,and using the improved SRGAN to reconstruct it,a new super resolution image can be obtained by collaboration between the two.Firstly,a soft tag function is proposed to replace the original fixed tag for DCGAN.Secondly,the void convolution residual block is introduced as the main structure of SRGAN discriminator,and the CBAM attention mechanism is added to realize the weight redistribution.Finally,adaptive average pooling is introduced in SRGAN discriminator to reduce the number of network parameters.The experimental results show that:using standard data sets AID and RSOD,by MobileNet V2 classification network test,DS-GAN data enhancement method is significantly improved compared with conventional enhancement method and DCGAN enhancement method.The accuracy of AID data set is improved by 8.01% and 9.49%,respectively.In the RSOD data set,the accuracy improved by 4.76% and 1.4%,respectively.