Visual transformer with stable prior and patch-level attention for single image dehazing

Jinzhe Liu, Heqiang Yuan, Zhiqiang Yuan, Li Liu, Bin Lu, Miao Yu

Abstract: Highlights•Transformer-based architecture to solve the dehaze ability decline of existing models in real scenarios.•A patch-level attention mechanism is proposed to correct and complement regional images.•Introduce an invariant prior to provide stable support for the dehazing process.•Design the obfuscation filling strategy to bridge the domain gap caused by the change of data distribution.•Superior results are achieved on both synthetic and real haze datasets, with domain adaptation and generalization capabilities.