Polarization imaging bi-attentional recursive residual super-resolution method

Abstract

Due to the limitation of the detector, the spatial resolution of the polarization image obtained by space-modulated full-polarization computed imaging is low. In general, there are a lot of valuable high-frequency components in low resolution space, but it is difficult to reconstruct high frequency information by deep learning method. Through the research of the previous super-resolution reconstruction method based on deep learning, it is found that it is difficult to obtain better lifting effect only by stacking residual blocks to build a deeper network. A polarization imaging bi-attention recursive residual super-resolution reconstruction method is proposed. In the network structure, bi-attention recursive residual group is used as the deep feature extraction module, and the hybrid bi-convolutional attention module contained in this module is used to adaptively learn image features of different channels and different transformation Spaces in the deep network. In order to provide the concentrated learning efficiency of high-frequency information, jump connections are introduced into the network structure, and the shallow feature extraction and reconstruction of images are completed by a convolution layer and a sub-pixel convolution algorithm respectively. The experimental verification is carried out under the actual imaging system and simulation data and compared with other methods in terms of visual effect and quantitative results. Experimental results verify that the proposed method can effectively reconstruct the contour structure and texture details of the detection target in subjective image quality and is superior to the comparison method in objective evaluation index.