Unsupervised Aerial Maritime Image Dehazing via Pooling Transformer and Multi-Granularity Attention Optimization

Fog degradation in aerial maritime images reduces contrast and detail, directly impairing operational efficiency. This paper proposes an unsupervised dehazing algorithm integrating Pooling Transformer and multi-granularity attention optimization. Based on the unsupervised multi-branch network with high-frequency enhancement (UME-Net), our improvements include: designing a lightweight Pooling Transformer Block (PTB) to replace residual blocks; developing a multi-granularity channel attention decoder (MFCA) surpassing conventional attention mechanisms; optimizing transmission-aware dynamic loss functions. Evaluations on our DIOR-seafog dataset demonstrate PSNR of 23.71 dB and SSIM of 0.919, achieving 4.18 % and 3.80 % improvements over baseline respectively, with superior visual quality compared to mainstream unsupervised methods. The solution effectively addresses image restoration in complex maritime environments.