Image-to-image Translation for Enlargement of Aircraft Skin Defect Datasets

With the size and diversity of a dataset being integral to the accuracy of aircraft skin defect detection models, it can be argued that the currently available small aircraft skin defect datasets need to be greatly increased. In this paper, we suggest employing image-to-image translation techniques, specifically the CycleGAN network combined with perceptual loss. By exploiting the CycleGAN network’s ability to generate images similar to those fed into the network, we aim to increase the size of an available aircraft skin damage dataset through the generation of synthetic images. The addition of perceptual loss further enhances the quality and realism of generated images. Through image evaluation with Fréchet Inception Score, we showcase how this method has the capacity to address the challenge posed by limited datasets for small aircraft skin damage.