Polarization-based image dehazing network with pseudo-3D convolution

In this study, we present a pseudo-3D convolutional feature fusion attention network specifically designed for polarization-based image dehazing. Within this network, we introduce a novel feature attention module based on the Pseudo-3D convolution structure, integrating spatial feature attention and polarization feature attention mechanisms. Through a differentiated weight assignment model, this module allocates varying attention to haze at different locations and thicknesses, and adopts diverse processing approaches for hazy images captured at different polarization angle channels. In addition, we introduce a basic block structure that combines local residual learning, an attention module, and an octaves convolutional residual module. This integration allows the network to disregard information from thin hazy regions and low-frequency details, focusing more on critical information, significantly enhancing network performance. Experimental results unequivocally demonstrate the state-of-the-art performance of our method on both synthetic and real-world hazy images.