Fig. 1. Channel attention based polarimetric dense network for underwater image polarization information restoration. (a) Schematic diagram of the network structure; (b) internal structure of the RDB; (c) structure of the CA block

Fig. 2. Structure of the content and style loss

Fig. 3. Experimental setup for underwater polarization image acquisition

Fig. 4. Demo of underwater polarization image datasets. (a) Pairs of original images captured by polarization camera; (b) polarization images of four angles after polarization down-sampling processing; (c) rearranged polarization images

Fig. 5. Results of restored intensity images and polarization informations of different network structures

Fig. 6. Comparison of image descattering and polarization information restoration effects by different methods

Fig. 7. Influence of the number of CA modules on the recovery of polarization information

Fig. 8. Influence of RDB parameters on network performance. (a) Number of blocks D; (b) number of convolutional layers C; (c) number of channels in each convolutional layer G

Fig. 9. Influence of loss function weights on restoration results. (a) Influence on restored light intensity image; (b) influence on recovered DoLP image; (c) influence on recovered AoP image