Robust Chinese License Plate Generation via Foreground Text and Background Separation

Abstract

To solve data scarcity, generating Chinese license plates with Generative Adversarial Network becomes an efficient solution. However, many previous methods are proposed to directly generate the whole license plate image, which causes the mutual interference of the foreground text and background. This way, it may cause unclear character strokes and an unreal sense of the overall image. To solve these problems, we propose a robust Chinese license plate generation method by separating the foreground text and background of the license plate to eliminate mutual interference. The proposed method can generate any Chinese license plate image while maintaining the precise character stroke and background of the real license plate. Specifically, we substitute the foreground text of the real license plate with the target text. To provide supervision data for text substitution, we propose to synthesize them via foreground text and background separation. Firstly, we erase the text of the real license plate to obtain the corresponding background image. Secondly, we extract the foreground text of another real license plate and merge it with the background obtained above. Qualitative and quantitative experiments verify that the license plates generated by our method are more homogeneous with the real license plates. Besides, we enhance license plate recognition performance with the generated license plates, which validates the effectiveness of our proposed method. Moreover, we release a generated dataset (https://github.com/ICIG2021-187) with 1,000 license plates for each province, including all 31 provinces of the Chinese mainland.