Differentiable Style Searching: An Online Automatic Data Augmentation Method

Aiming at the problems of offline image transformation and limited search space in existing data enhancement methods, an online automatic data enhancement (ODA) method is proposed. The core of ODA is a micro-style search module, which can perceive the current data enhancement required by the task model by directly transmitting training losses, and generate more difficult stylized pictures online to expand the training set by means of anti-search, so as to effectively help the model complete the generalization on a variety of unknown styles. On the cross-domain image classification task of MNIST, MNIST-M, SVHN and USPS data sets, as well as the cross-domain scene semantic segmentation task of Cityscapes and GTA5 data sets, compared with other five typical data enhancement methods, ODA method can improve classification accuracy by at least 2% under Acc metric and 3% to 7% under mIoU metric of semantic segmentation task, which proves that ODA extends the search space of traditional automatic data enhancement methods in the direction of image style and enhances the generalization ability of network.