Image Augmentation based on Cross Domain Image Style Transfer

Data is the cornerstone of various existing deep learning methods. The basic paradigm of existing deep learning methods is to build a deep neural network model, and make the model parameters fit the internal representation of datasets through back-propagation and gradient descent. Data augmentation is a technology that uses existing data in the small datasets to expand the overall datasets scale. Image style transfer is a research hot spot in the field of computer vision. It can endow the image to different styles while keeping the image content unchanged. This paper proposes an image augmentation method based on cross domain image style transfer, it is divided into two parts: cross domain image style transfer and smooth. In the cross domain style transfer stage, the input image is transfer to the domain of the style image, and the smooth stage can eliminate the artifacts in the generated image. Experiments show that this method is effective, it can expand large-scale datasets on the basis of existing data, and enhance the ability of neural network model to obtain better performance in the train process.