Hydrological Image Data Enhancement Based on Generative Adversarial Network and Image Pyramid Mechanism

With the continuous development of intelligent water conservancy and the wide application of computer vision technology in the field of hydrology, the problem of lack of hydrological images and poor quality of hydrological images has emerged. This problem limits the performance improvement of deep learning related research models in hydrology, such as semantic segmentation and target detection, etc. In this paper, a deep learning model based on generative adversarial networks and image pyramids, LapSLC-GAN, is proposed for image generation of hydrological images in the hydrological domain. This model accomplishes image data enhancement, improves the quality of the synthetic images, and solves the problem of unclear image details. In this paper, we have improved the generator. Firstly, we propose a new downsampling module based on the semantic segmentation map to fully extract features and exploit them. Secondly, the model integrates an image pyramid mechanism to enhance the emphasis on high-frequency information and improve semantic loss. This is achieved by overlapping the semantic segmentation map and the image pyramid. On the homemade hydrological image dataset, this paper designs experiments to demonstrate the effectiveness of the downsampling module and LapSLC module. It is also compared with other GAN models. The numerical results show that the model proposed in this paper can improve the accuracy of the segmentation model by about 8% compared with the better-performing models at this stage.