HF-DehazeNet: Hybrid Features-Based Network for Image Dehazing

Image dehazing is an important low-level computer vision issue that reconstructs latent haze-free images form haze images. In recent years, many end-to-end deep learning-based dehazing methods manifest a better potential of reconstruction performance than traditional atmospheric scattering model-based methods. In this paper, we propose a hybrid features-based dehazing network, termed as HF-DehazeNet, which combines convolution and self-attention operations to extract local feature and global representations of hazed image, respectively. Furthermore, a fusion block is used to fuse all the feature maps from different channels. To demonstrate the effectiveness of the proposed network, we compare several typical approaches. Experiments show that HF-DehazeNet strengthens the ability of representation learning, consequently, it outperforms the previous methods on the most frequently used RESIDE dataset, especially for the hazy images with rich texture details and sky regions.