A Novel U-Shaped Hybrid Network for Single Image Dehazing

Abstract

Image dehazing is a challenging problem due to its ill-posed parameter estimation. Despite the significant success of Convolutional Neural Network (CNNs), the inherent locality of CNNs remains a bottleneck for dehazing performance. Though Transformers mitigate the shortcomings of CNNs and have demonstrated promising performance in high-level vision task, the inherent computational complexity makes them infeasible for low-level vision task. In this work, an efficient U-shaped Convolution and Transformer hybrid network, called UCPformer, is proposed. Specifically, Channel Enhanced Transformer (CET) and Efficient Pixel Enhanced Transformer (EPET) is designed in this paper for efficient encoding and decoding of hazy image features. The CET inherits the local representation capability of CNN and general architecture of Transformer, extracting local information efficiently and treating different channels unequally. The EPET inherits the global context modeling capability of Transformer, treating different pixels unequally with linear complexity. Experiments demonstrate the proposed UCPformer achieve superior performance against other dehazing methods.