Coarse-to-Fine Perception Transformer with Frequency Awareness for Single Image Dehazing
Wangqianjin WU, Lihan Tong, Zhiqi Lin, Yendo Hu, Erkang Chen · IEICE Transactions on Information and Systems · 2026
Transformer-based dehazing models often struggle with limited perception granularity and overlook high-frequency details. To address these issues, we propose a Coarse-to-Fine Perception Transformer Network with Frequency Awareness (CFPT-Net). CFPT-Net introduces a three-level sparse attention mechanism, operating across coarse (4×4), medium (2×2), and fine (1×1) region scales to capture multiscale token affinity. On the coarse and medium scales, we apply TopK sparse attention to model global dependencies efficiently. On the fine scale, a windowed attention strategy with ReLU2 activation is used to reduce computational cost while improving the representation of local features. We also incorporate frequency-domain information in the feedforward network of the transformer block to preserve image details lost in hazy conditions. Experimental results demonstrate that CFPT-Net achieves competitive and favorable dehazing performance both quantitatively and visually.