Truly Generative Data Augmentation for Image Segmentation - Case of Cloud Images

Mayank Jain, Soumyabrata Dev

Keywords: GAN, data augmentation, sky, cloud, image segmentation

TL;DR: This paper presents an organic, reliable and generative approach for data augmentation towards sky/cloud image segmentation.

Abstract: Supervised learning frameworks frequently rely on semantic image segmentation, which necessitates a substantial amount of annotated data. Existing methodologies for data augmentation either employ image transformations that are limited by the cardinality of the original dataset or employ generative augmentation techniques that introduce pixel categorization errors. This paper presents an innovative approach for "truly" generative data augmentation for image segmentation, specifically in the context of sky/cloud images. The proposed method involves separate generation of the background clear sky image and the foreground cloud masks using two separate DCGANs, which are subsequently merged to produce augmented images. This organic approach enhances the quality of generated images while preserving accurate pixel categorization. The proposed approach is finally noted to improve the robustness of the sky/cloud image segmentation models.