Fig. 1. Proposed network structure

Fig. 2. Structure of PMAB. (a) Module 1; (b) module 2; (c) module 3

Fig. 3. Receptive field of hybrid dilated convolution contrast

Fig. 4. Attention module structure

Fig. 5. Attention module feature map visualization. (a) Hazy image; (b) attention module output feature map

Fig. 6. Selective feature fusion block structure

Fig. 7. Comparison of experimental results of synthetic hazy images by different algorithms. (a) Hazy images; (b) DCP; (c) MSCNN; (d) AOD-NET; (e) AEDGAN; (f) URNET; (g) HR-Dehazer; (h) PCFAN; (i) MADN; (j) proposed algorithm

Fig. 8. Comparison of experimental results of different algorithms on real hazy images. (a) Hazy images; (b) DCP; (c) MSCNN; (d) AOD-NET; (e) AEDGAN; (f) URNET; (g) HR-Dehazer; (h) PCFAN; (i) MADN; (j) proposed algorithm

Fig. 9. Comparison chart of PSNR and SSIM results of different algorithms on SOTS dataset. (a) PSNR; (b) SSIM

Fig. 10. Comparison chart of Entropy,FADE,VIF,AG results of different algorithms on HSTS dataset. (a) Entropy; (b) FADE; (c) VIF; (d) AG

Fig. 11. Comparison diagrams of dehazing effects of different module ablation experiments. (a) Hazy image; (b) residual block; (c) attention residual block; (d) parallel multi scale mapping block; (e) PMAB; (f) PMAB+up-down sampling fudion; (g) PMAB+up-down sampling fudion+SFFB; (h) hybrid dilated convolution