Image Generation Framework for Unbalanced License Plate Data Set

Deep learning based methods have achieved promising performance in the image fields, but they are significantly sensitive to the distribution of data. To obtain balanced data, the acquisition of annotation data is very time-consuming and laborious. Generative Adversarial Networks have made a lot of progress in data generation but may degrade greatly in the unbalanced data set. In this paper, we propose an image generation framework to generate photo-realistic, various, and balanced images of text. Specifically, we add the module for training unpaired images (U-module) and target selector to our framework, and the target selector uses text string to select images in the extended real images, which contain real images and the generated images by the U-module, in addition, global generator and local enhancer network are applied to improve the quality of the generated images. We demonstrate our method on the Chinese license plates, and the unbalance of the license plate data set is shown in the provincial category and the special license plates. The Inception Score and the FID Score are used as metrics to validate our method. The experimental results show that the Inception Score of the generated images is close to that of the real images, and our method achieves the lower FID Score than the state-of-the-art. In the SYSU-ITS dataset, the accuracy of license plate recognition has been largely improved, especially for the provinces with no or few real samples.