Dual-domain multi-scale feature extraction for image dehazing

Image dehazing is a typical task in the field of computer vision. Its aim is to restore clear images from hazy images, remarkable research accomplishments have been achieved in image dehazing approaches based on convolutional neural networks. However, existing research still has two main limitations: the exclusive use of conventional or large convolution kernels is insufficient to extract multi-scale features from images, making it challenging to balance the structure and details in the reconstructed images; existing methods neglect the critical information contained in the frequency domain, resulting in indistinct color transitions in the image. This paper proposes an image dehazing network based on Dual-Domain Multi-Scale Feature Extraction (DMFE-Net) to address these issues. Specifically, the multi-scale dilated convolution kernels of DMFE-Net expand the receptive field in the spatial domain to capture the multi-scale structural features of the image and constitute the Structure Feature Extraction Block (SFEB). The texture detail features of the image are supplemented by conventional convolutional kernels and constitute a Texture Feature Extraction Block (TFEB), which works together with the SFEB to capture the multi-scale features of the image. In the frequency domain, Frequency Domain Feature Enhancement Block (FDEB) is devised to learn the amplitude differences at the same frequency between the hazy and the clear images. Moreover, a step-by-step fusion strategy is designed to integrate features extracted by different types of blocks. Finally, DMFE-Net is structured in a U-Net-like architecture. Compared with previous state-of-the-art methods, DMFE-Net successfully restored more image details while ensuring structural integrity, with a more pronounced color transition. DMFE-Net achieved a maximum PSNR of 40.32 dB and SSIM of 0.993 on the SOTS-indoor dataset. Extensive experiments on multiple datasets validated the efficacy of the proposed method. Code release is available at https://github.com/chingguo/DMFE-Net.