<p>Currently, significant progress has been made in conventional-size image dehazing technology, but restoring ultra-high-definition images is still a challenging task. Existing ultra-high-definition image dehazing methods usually cut down the resolution of the image to adapt to the network architecture, thereby avoiding large-scale convolution operations. However, traditional downsampling methods may result in the loss of a large amount of high-resolution information. To solve this problem, this paper proposes a multi-scale ultra-high-definition image dehazing network. Using the Laplacian pyramid to extract multiple features with different resolutions from high to low and incorporating convolution blocks and SimAM modules to improve feature expression. To achieve the restoration of high-quality haze-free images, the ResNet block in the HRNet network is adjusted to MSBlock with parallel multi-scale convolution kernels, thereby strengthening the model’s receptive field. In addition, this article uses depthwise separable convolution to reduce model computational effort, while ensuring the network dehazing speed. A large number of skip connections are introduced to retain multi-scale information, and multi-resolution subnetworks are connected in parallel. By repeatedly fusing multiple scales, we obtain rich high-resolution features, thereby generating clear haze-free images. This paper quantitatively and qualitatively compares the proposed model with several existing state-of-the-art algorithms on the 4KID, I-Haze, and O-Hzae datasets. Experimental results show that the proposed method can not only maintain real-time processing speed but also achieve a high peak signal-to-noise ratio and structural similarity.</p>