Underwater polarization de-scattering method using residual dense block and depth-wise convolution

We propose an underwater polarization de-scattering method based on deep learning and an improved U-net to cope with the imaging challenges in underwater turbid environments. Firstly, we present a feature extraction and fusion module based on residual dense block and depth-wise convolution (RDD) to achieve efficient feature extraction and local information encoding. Second, we design a down-sampling module with low computational complexity to preserve richer features, and the up-sampling module is optimized using transposed convolution. To validate our method, we constructed underwater polarization datasets with different turbidity and targets, and compared it with existing de-scattering methods. Experimental results demonstrate that our method significantly outperforms existing underwater de-scattering imaging approaches in terms of restored image quality and detail preservation. In particular, our method shows robustness in different underwater turbidity environments, which provides a new solution for underwater clarity imaging.