Integrated multi-branch neural network dehazing algorithm

Recently, the field of image dehazing has achieved rapid development. End-to-end deep learning algorithms achieve good performance. However, we observe that many algorithms perform poorly on the task of image dehazing in high-resolution complex environments. In this paper, aiming at the problems of detail loss, edge blurring and regional haze legacy in the process of dehazing non-uniform haze images with high resolution by existing algorithms, an end-to-end dehazing model integrating multiple neural networks is proposed. To be specific, we introduce a multiple neural network to deal with the aforementioned problems. Our proposed structure consists of three branches: the encoder-decoder network is constructed by transfer learning to reconstruct haze-free features; Full-resolution residual extraction network helps to avoid loss of details in downsampling layers; Swin-transformer scale extraction network makes the network receptive field richer. Then, their different features are then mapped through a learnable fusion tail. Extensive experimental results on real datasets show that our proposed ensemble multi-branch dehazing model has better recovery ability and good generalization ability on multiple datasets.

I. Introduction

As suspended particles such as water vapor, small particles and dust converge when the climate decreases, the color of the image will turn gray and the whole image will be covered with layers of gray haze. So we need to use image dehazing technology to improve the image clarity.