A dual-branch residual network for inhomogeneous dehazing

Image dehazing has now gained the dominant popularity in the field of image processing, particularly in inhomogeneous scene. Recent years have witnessed great progress in handling homogeneous dehazing problems. Due to the unknown haze distribution of the real world, it is extremely intractable to offer a clear view of the observed scene in limited inhomogeneous datasets. Furthermore, it is impossible to completely avoid artifacts of color distortion, over -enhancement, halo, and blur errors in order to provide reliable and stable results. To this end, we propose an end -to -end Dual -branch Residual Network (DBRN) for inhomogeneous dehazing that is composed of Hybrid Feature Subnet (HFS) and Attention Feature Fusion Subnet (AFFS). HFS explores high -quality global and local hazy features with long-range spatial fusion at different scales in an encoder-decoder manner, while AFFS makes artifact removal possible with a stack of Residual Convolution Attention Module (RCAM). Besides, the joint loss function aims to ensure that the recovered image is close to the ground -truth in the aspects of texture, color, structure index, and so on. Through this design, the model exhibits robustness in inhomogeneous hazy scenes, enabling high -quality visual restoration in scenes with varying haze densities. Extensive experimental results demonstrate that the proposed model performs favorably against the state-of-the-art methods on synthetic datasets and real -world hazy images. In addition, ablation studies are carried out to demonstrate the effectiveness of each component. The source code of the proposed method is available at https://github.com/jing-1196/DBRN/.