基于改进DCGAN算法的遥感数据集增广方法

Augmentation Technology of Remote Sensing Dataset Based on Improved DCGAN Algorithm

遥感数据集规模是深度学习目标检测算法性能的关键,如何利用少量数据生成大量标注图像成为当前的研究热点。针对这一问题,结合二次掩模技术,提出一种基于改进DCGAN算法的遥感数据集增广方法,自定义目标个数与位置,实现图像与标签的扩增,解决了基于GAN图像增广算法中无对应标签生成的问题。同时,针对DCGAN算法生成图像质量不高的问题,提出多尺度特征融合技术,优化DCGAN算法,提升图像质量。实验表明,在MNIST和PlANE两种数据集上,改进DCGAN算法生成的图像质量与图像多样性均优于DCGAN算法;在利用Tiny-YoloV2算法设计的验证实验中发现,所提算法增广的数据集,检测AP值高达85.45%,相对未增广算法与传统增广方法,AP值分别提高了16.05%和2.88%,验证了算法的有效性。

The scale of the remote sensing dataset is a crucial factor to object detection algorithm’s performance based on deep learning technology.How to generate large number of labeled images by using a small amount of data has become a hot research topic.To solve the problem,we propose an augmenting method of remote sensing dataset based on improved DCGAN algorithm using the secondary mask technology.In addition,the algorithm proposed in this paper can realize the amplification of images and labels by determining the number and location of targets to be generated,to solve the problem of no corresponding label generation in GAN-based image augmentation algorithm.Moreover,a multi-scale feature fusion technique for optimizing DCGAN algorithm is proposed to solve the problem of poor image quality generated by DCGAN algorithm.Experiments show that the improved DCGAN algorithm is superior to the DCGAN algorithm on both MNIST and PlANE datasets in terms of image quality and image diversity.The AP value of the dataset expanded based on the method we arranged is up to 84.45%in the design experiment using the Tiny-YoloV2 algorithm.Compared with the unaugmented algorithm and the traditional augmented method,the AP value of the method adopted in this paper is increased by 16.05%and 2.88%respectively,which fully verifies the effectiveness of the the technical scheme designed in this paper.