Efficient Multi-Scale Transformer with Convolutional Attention for High-Definition Image Dehazing

Image dehazing aims to recover clear images from hazy inputs, crucial for advanced computer vision tasks. While convolutional neural networks (CNNs) have traditionally been used, transformer-based methods have recently demonstrated superior performance in capturing global information. However, their computational complexity scales quadratically with image resolution, posing challenges for high-resolution tasks. To address these limitations, we propose Mulsormer, a novel transformer-based dehazing model tailored for high-definition images. Mulsormer incorporates Multi-Head Convolutional Attention (MHCA) and a Gated Feed-Forward Network (GFFN) to reduce computational complexity and control critical information flow, respectively. A multi-scale loss function combining perceptual and patch-based losses is utilized to improve detail restoration. Extensive experiments on the SOTS and RTTS datasets show that Mulsormer achieves state-of-the-art performance, with a PSNR of 37.49 dB and an SSIM of 0.993 on the SOTS Indoor dataset, significantly outperforming previous methods. Our approach demonstrates the potential of transformer-based models for efficient high-definition image dehazing. Additionally, our code is publicly available at https://github.com/Wang1666570/Efficient-Multi-Scale-Transformer.