Controlling the Quality of GAN-Based Generated Images for Predictions Tasks

Abstract

Recently, Generative Adversarial Networks (GANs) have been widely applied for data augmentation given limited datasets. The state of the art is dominated by measures evaluating the quality of the generated images, that are typically all added to the training dataset. There is however no control of the generated data, in terms of the compromise between diversity and closeness to the original data, and this is our work’s focus. Our study concerns the prediction of soil moisture dissipation rates from synthetic aerial images using a CNN regressor. CNNs, however, require large datasets to successfully train them. To this end, we apply and compare two Generative Adversarial Networks (GANs) models: (1) Deep Convolutional Neural Network (DCGAN) and (2) Bidirectional Generative Adversarial Network (BiGAN), to generate fake images. We propose a novel approach that consists of studying which generated images to include into the augmented dataset. We consider a various number of images, selected for training according to their realistic character, based on the discriminator loss. The results show that, using our approach, the CNN trained on the augmented dataset generated by BiGAN and DCGAN allows a significant relative decrease of the Mean Absolute Error w.r.t the CNN trained on the original dataset. We believe that our approach can be generalized to any Generative Adversarial Network model.

The authors thank Campus France and Morocco CNRST for the financial support of this research, and Telecom SudParis, and Univ. Mohammed V during the different phases of this study. This work was also supported in part by the National Natural Science Foundation of China under Grant 61976030 and the funds for creative research groups of Chongqing Municipal Education Commission under Grant CXQT21034.