Water Gauge Image Augmentation based on Generative Adversarial Network

Abstract

Water level monitoring based on water gauge is a very widely used way because of its cheapness and portability. However, the insufficiency and low quality of water gauge images restrict the performance of water level measuring task based on deep learning methods such as object detection and semantic segmentation. In this article, we proposed a generative adversarial network (GAN) named Contextual adjustment GAN (CA-GAN) for data augmentation of water gauge images obtained from Wuyuan, Jiangxi Province in China, i.e. CA-GAN can generate high-quality images containing various scales and types water gauge, which provide image data for application such as deep-learning based water level measuring method. First, a improved downsampling module is designed with the help of segmentation map for the semantic activation modulation. Then, the Unet++ structure with the improved downsampling module is applied in the generator. Finally, to modulate the semantic relationship, a contextual adjustment scheme is de-signed between adjacent layers. This article conducts detailed experiments, proving that the improved downsampling module contributes to the maintenance of semantic information of water gauge images. It is illustrated that the water gauge images generated by CA-GAN have higher quality comparing with other three GAN models. And our method is expected to promote the water level measurement and hydrological monitoring application development.