KA-DMSF: kernel attention and dynamic multishape filtering for real-world image dehazing

This paper proposes an efficient image dehazing network named KA-DMSF. By introducing depth-wise separable convolutions and channel attention mechanisms, this work significantly reduces the number of parameters and computational cost while maintaining strong feature representation capabilities. Additionally, we design a composite loss function that combines multi-scale L1 loss and frequency domain (FFT) loss, enabling the model to simultaneously recover detailed textures and global structural information in images. Experimental results on multiple public datasets demonstrate that the proposed method achieves state-of-the-art performance in both subjective visual quality and objective metrics (such as PSNR and SSIM), while also providing faster inference speed. This offers an efficient and practical solution for image dehazing tasks.