Dual-channel encoding model in spatial and frequency domains for underwater polarimetric imaging

In recent years, learning-based underwater polarimetric imaging models have undergone rapid expansion. Unfortunately, the majority of learning-based models have limitations in feature extraction and fail to make full use of frequency domain features. To further improve restoration capability, we present a dual-channel encoding model in the spatial and frequency domains for underwater polarimetric imaging. First, to effectively restore the high- and low-frequency features of hazy polarization images, we utilize two subnetworks to decompose the images into high- and low-frequency components, enabling the network to recover the hazy polarization images on the two feature components. Specifically, we employ a lightweight encoder–decoder architecture to restore the low-frequency feature components. Meanwhile, for the high-frequency feature components, we introduce a well-designed high-frequency aggregation component, which recovers the high-frequency features of the current region by referring to neighboring feature distributions that are not completely corrupted by backscattered light. Second, we introduce an additional spatial domain network integrating an active polarization imaging model proposed in our previous work to directly restore spatial features. Lastly, the results from the frequency and spatial domain networks are fused to reconstruct clear images. Experimental results on the established underwater polarization dataset verify that our method, to our knowledge, outperforms other advanced methods.