Residual Hybrid Attention Network for Single Image Dehazing

Abstract

In hazy weather, there is an excess of gaseous and liquid particles in the air. Due to the scattering and absorption of these particles, images acquired by imaging devices typically exhibit degradation such as reduced contrast, brightness, and visibility, which affects the effectiveness of subsequent intelligent systems. To address this degradation, a large number of deep learning-based algorithms have been proposed to effectively remove haze. Recent studies have improved the feature learning capability of models with multi-scale fusion strategies and multi-branch networks. These complex structures improve the accuracy of haze removal, but come with some side effects, such as a larger number of parameters, increased memory consumption, and slower running speed. On the contrary, a single sequential network has a simple structure and a relatively small number of parameters. Furthermore, sequential operations can avoid the additional memory consumption caused by the large number of hierarchical features residing in memory. Different schemes attempt to use full-resolution and deeper networks to obtain better haze-free images, but also suffer from issues with parameters, memory, and running speed. To balance model performance and memory consumption, we propose a single image dehazing network based on encoder-decoder structure that integrates residual learning and attention mechanisms. The resolutions of feature maps are reduced by down-sampling, and the feature refinement module gradually recovers haze-free images at low resolution. Moreover, the receptive field of the model is increased by the large kernel design, which can capture global features and improve feature learning. The hybrid attention mechanism adaptively adjusts the spatial and channel weights so that the network will pay more attention to task-relevant objects and regions for accurate haze removal. Experiments on synthetic and real-world datasets demonstrate the effectiveness of the proposed algorithm.