Efficient ultra-high-definition image dehazing based on multi-feature self-attention

Transformers have demonstrated excellent performance in single-image dehazing tasks, with steadily improving dehazing performance. However, when directly applied to ultra-high-definition (UHD) image dehazing, the high resolution of UHD images significantly increases inference time. Additionally, their global modeling nature tends to overlook local details. To address these issues, this paper proposes an efficient UHD image dehazing network called CSDehazeNet based on multi-feature self-attention. This network introduces multi-scale convolutions within ConvNeXt blocks to extract haze-related features. After feature dimensionality reduction, these features are fed into the Transformer to reduce computational complexity. Meanwhile, in the Transformer, depthwise separable convolution, deformable convolution, and linear transformation are adopted in three parallel branches to generate query vectors (Q). This guides the self-attention mechanism to focus on local structural information, enhances the expression of local details, thereby improving the restoration of local details. Furthermore, this paper proposes a spatial statistical normalization method based on LayerNorm, which is introduced into both the ConvNeXt and Transformer modules to enhance the network’s robustness in complex hazy scenarios. Finally, experimental results on three datasets (4KID, O-HAZE, and 8KDehaze) show that the proposed method not only achieves higher-quality image restoration but also improves dehazing efficiency. On the 4KID dataset, it achieves a PSNR of 31.23 dB with only 1.30M parameters, and an inference time of merely 0.2 seconds. Code is available at https://github.com/wangkunxing/CSDehaze.