基于改进DCGAN的对地观测图像生成方法
Earth observation image generation method based on improved DCGAN
黄丹丹 1汪梅 1张永高 1施俊杰 1张岩 2李远成1
作者信息
- 1. 西安科技大学计算机科学与技术学院,陕西西安 710054
- 2. 西安科技大学电气与控制工程学院,陕西西安 710054
- 折叠
摘要
为了研究无人机对地观测图像样本的平衡性,提高对地观测在深度学习中的应用,采用图像生成方法对无人机对地观测图像进行大量生成;针对图像生成模型在训练时出现的稳定性和生成图像的质量问题,提出一种基于改进DCGAN的对地观测图像生成方法.首先在DCGAN的生成器和判别器的网络结构中增加批处理层,然后将判别器的优化器改进为随机梯度下降,且生成器的优化器采用自适应学习率,最后改进模型的损失函数.结果表明:改进后的DCGAN网络模型生成的数据与原始数据的统计特征相似,模型性能良好,相比于其他的GAN衍生模型,改进后的DCGAN模型更具有稳定性,在训练过程中未出现模式崩塌的现象,模型生成图像的FID分数值为4.631,比原始DCGAN模型低2.409,该方法生成的图像质量更好,更加适用大规模的对地观测图像数据的生成.
Abstract
In order to study the balance of UAV ground observation image samples and improve the ap-plication of ground observation in deep learning,an image generation method is used to generate a large number of UAV ground observation images.For the stability of the image generation model during train-ing and the quality of the generated images,a ground observation image generation method based on improved DCGAN is proposed.Firstly,a batch processing layer is added to the network structure of the generator and discriminator of DC GAN;secondly,the optimizer of the discriminator is improved to sto-chastic gradient descent and the optimizer of the generator adopts adaptive learning rate,and finally,the loss function of the model is improved.The experimental results show that the data generated by the im-proved DCGAN network model is similar to the original data in terms of statistical characteristics,and the model performance is good.Compared with other GAN-derived models,the improved DCGAN model is more stable,and there is no pattern collapse during the training process,and the FID score value of the model-generated image is 4.631,which is 2.409％lower than that of the original DCGAN model,indicating that the quality of the image generated by the proposed method is very high.The FID score of the model generated image is 4.631,which is 2.409 lower than the original DCGAN model,indica-ting that the proposed method generates better the images in quality and is more suitable for large-scale Earth observation image data generation.
关键词
对地观测/深度卷积生成对抗网络/深度学习/图像生成Key words
earth observation/deep convolutional generative adversarial networks/deep learning/image production引用本文复制引用
基金项目
国家重大专项项目(2022ZD0119005)
西安市重点产业链核心技术攻关项目(23ZDCYJSGG0025-2022)
陕西省自然科学基金项目(2023JC-YBMS-539)
出版年
2024
国家科技期刊平台
NSTL
万方数据