Efficient Re-Parameterization Residual Attention Network for Nonhomogeneous Image Dehazing

Real-world nonhomogeneous haze brings challenges to image restoration. More efforts are needed remove dense and thin simultaneously efficiently. However, most existing dehazing methods do not pay attention the complex distributions of usually suffer from a low runtime speed. To tackle such problems, we present an efficient re-parameterization residual network (RRA-Net), whose design has three key aspects. Firstly, propose training-time multi-branch block (MRAB), where multi-scale convolutions in different branches cope with nonuniformity converted into single-path convolution during inference. It also features local learning improved spatial channel attention, allowing be attended differently. Secondly, our lightweight structure cascades six MRABs followed by long skip connection fusion tail. Overall, RRA-Net only about 0.3M parameters. Thirdly, two new loss functions, namely Laplace pyramid color attenuation loss, help train recover details colors. The experimental results show that proposed performs favorably against state-of-the-art on real-world datasets, including both homogeneous haze. A comparison under same hardware setup demonstrates superior efficiency network.