A GAN Based Data Augmentation Method for Road Pothole Detection

Abstract

The detection accuracy of road potholes is not high due to the small number of positive samples. In order to expand the training dataset of the detection network and improve its detection accuracy, this paper proposes a road pothole data augmentation method combining generative adversarial network and image fusion technology. In this method, the clear forged pothole images with different morphometry are generated separately through SinGAN network, and the pothole image and road image are synthesized by Poisson image fusion. A mask image generation method for Poisson image fusion is also presented to further improve the edge smoothness of the fused part. The results of experiments have shown that the image samples generated by this method can significantly improve the accuracy of the road pothole detection network, and verified the effectiveness of this method.