Data Augmentation of Aerial Traffic Images Based on Optimal Transport Theory

Due to the issue of UAV perspective, aerial traffic images often fail to achieve full-area full-angle coverage. In order to solve the problem of scarce and low-quality samples of aerial traffic images, data augmentation using image generation models has become a popular method for integrating advanced information technology in the transportation field. Currently, images generated by image generation models suffer from issues such as low image quality and difficulty in generating diverse samples. Therefore, to address these challenges, this paper proposes a new image generation model: the Autoencoder-Optimal transport model. This paper explains the image generation task from a geometric perspective, which involves two steps: manifold learning and probability distribution transformation. Firstly, an autoencoder is constructed to learn the underlying manifold. Secondly, a semi-discrete optimal transport network is established for probability distribution transformation. Finally, these two parts are combined to form an Autoencoder-Optimal transport model. Experimental results using aerial traffic images are analyzed to demonstrate the model's ability to generate realistic aerial traffic images.