Ground-Based Cloud Images Recognition Based on GAN and PCANet

Abstract

In order to improve the recognition accuracy of ground-based cloud image, a new algorithm based on deep learning is proposed. In our approach, a large number of cloud images are generated by Generative Adversarial Networks first. Then, based on the original and these generate cloud images, the deep features of cloud images are extracted automatically by multi-layer automatic sensing feature network, which increase the features description ability effectively. Finally, the Support Vector Machine (SVM) classifier is trained and the cloud image recognition is completed. Comparing with the methods such as gray level co-occurrence matrix (GLCM) and PCA with original database only, our approach combines the advantages of both GAN and PCANet, and the experiment results shows that the accuracy of cloud image recognition is significantly improved.