A U-Net Based Image Dehazing Network with Parallel Multi-Scale CNN and Vision Transformer

Image dehazing aims to restore haze-free images with both high fidelity and structural-textural integrity from atmospherically degraded images. This paper proposes a dualbranch multi-scale CNN-Vision Transformer (ViT) parallel architecture for image dehazing. The framework effectively combines the complementary strengths of CNN for local feature extraction and ViT for global context modeling. Specifically, the CNN branch employs a multi-scale feature fusion strategy. By deploying convolution kernels of different sizes, it purposefully extracts local details at small scales, effectively addressing the differential blurring of features at various scales caused by haze and enhancing the recovery of local details in images. The ViT branch is used to maintain global modeling capabilities, ensuring that the dehazed images remain structurally consistent. Experimental results demonstrate that the proposed method achieves an average PSNR exceeding 30 dB and an SSIM above 0.96 on several synthetic datasets, exhibiting favorable qualitative and quantitative evaluation performance.