CATIT: Cross-Adaptive Transformer for Road Image Translation

In the high-resolution image style transfer task of road traffic scenes, how to fully transfer style information while retaining the original content structure information is a challenging problem. In this paper, a novel Transformer-based generative adversarial network for high-resolution un-paired image translation is proposed. Firstly, we design a style transformation module based on cross adaptive Transformer, which dynamically adjusts the content features to achieve statistical alignment between content features and the target style. Meanwhile, an image frequency-domain enhancement module is designed based on cross-attention, which fuses the global information of the low-frequency style with the local details of the high-frequency content information. The detailed texture is then enhanced while the image style consistency is maintained. Furthermore, we design a threshold-guided negative sample screening strategy based on contrastive learning, which can improves the model's transfer effect. The experimental results well demonstrate the effectiveness of the proposed method.