Imbalanced Adversarial Learning for Weather Image Generation and Classification

Abstract

Image classification is a fundamental and significant task which has many potential computer vision applications, such as classification with data augmentation. Weather classification is such a case of imbalanced distribution of labels, because some types of weather, such as rain and snow, are relatively rare compared to sunny and haze days, and haze days are relatively infrequent category compared to sunny days. In this paper, we propose a data augmentation method using generative adversarial networks (GAN). It can supplement and accomplish the diversity of image data. Specifically, we designed a framework, using a deep convolution generative adversarial networks (DCGAN) as a generator to generate images to balance the imbalanced data, and use the Convolutional Neural Network (CNN) model as a classifier to verify the classification results. In order to verify the performance of DCGAN, we also propose an evaluation method on three benchmark datasets as a comparative experiment. The empirical results demonstrate that Using DCGAN, high-quality weather pictures can be generated on weather data sets. Additionally, after using DCGAN based data augmentation technology, our classification accuracy has improved.