AURA-Net: Image Dehazing Algorithm with Multi-scale Detail Enhancement
Chaoyu Kuai, Li Liang · 2024
When processing images, the current dehazing algorithm has problems such as color distortion, loss of texture details, and overall dimness of the image. This paper proposes a new end-to-end network model AURA-Net. First, an enhanced feature residual module (EFR Module) is designed to address the common color distortion and texture detail loss problems in dehazing algorithms. By introducing the residual structure, the EFR module can effectively enhance the expression ability of image features, thereby better retaining the detailed information of the image, reducing color distortion, and improving the quality of the dehazed image. Secondly, a multi-scale spatial attention deblurring module (MSAD Module) is designed. This module combines spatial attention and channel attention mechanisms to focus on blurred areas and important channels in the image. Through multi-scale feature extraction, the MSAD module can more accurately restore the details in the image. Finally, the mean square error (MSE) loss function is redesigned to help the network converge faster. Experimental results show that the dehazed image restored by the algorithm proposed in this paper is not only subjectively natural in color and clear in details, but also outperforms the existing mainstream algorithms in objective indicators.