High-performance underwater polarization imaging based on a dual-scale parallel network with TaylorFormer-CNN

Abstract

Underwater polarization imaging (UPI) remains challenging due to the unique physical properties of underwater environments, where images typically suffer from severe degradations. Despite the remarkable progress of deep learning-based methods, existing methods primarily rely on local spatial features from polarization images, failing to capture global context information. Here, we propose a dual-scale parallel network with TaylorFormer-CNN (DSPN-TC) for high-performance UPI. Specifically, in the Transformer branch, we introduce TaylorFormer, an efficient variant that approximates the softmax attention through a first-order Taylor expansion and leverages the associative property of matrix multiplication. For the convolutional neural network (CNN) branch, we design a deformable dense block to better adapt to underwater targets of diverse shapes and scales. Moreover, a dual-branch convolutional block is integrated into the encoders and decoders of both branches to generate dual-scale features with different receptive fields and semantic levels, thereby enriching the diversity of feature representations. Based on a constructed UPI experimental setup, we built a new UPI dataset. Experimental results demonstrate that the DSPN-TC exhibits strong adaptability and robustness in complex underwater environments, achieving state-of-the-art performance across multiple evaluation metrics.