In polarization imaging detection, the difference and change of physical properties of aerosols or detection targets are reflected by polarization characteristics. The high-dimensional polarization characteristics effectively improve the contrast between the target and the background, thereby laying the foundation for realizing the inversion of the target's spatial structure. This feature can enhance the recognition effect of the target in the cluttered background. Affected by the imaging distance and atmospheric interference, the limit resolution of the image projected on the focal plane is greatly reduced (much smaller than the optical system diffraction limit resolution), resulting in a lower spatial resolution of the polarized image. On the other hand, the spatial resolution of the polarized image is limited by the number of detector pixel. High resolution images are of great significance and value to the accuracy of target detection. For this reason, without replacing the hardware imaging system, the super-resolution reconstruction method is usually adopted. This method is a common technical means in image processing and practical engineering applications, and it is also a hot research issue of underlying computer vision. The existing image super-resolution algorithms have some problems, such as low utilization of feature information, large amount of parameters, blurred image reconstruction details and so on. The input features of low-resolution images contain rich low-frequency information, which is treated equally in different channels. In computational imaging process using deep learning, image mapping function solution space is very large, it is difficult to generate detailed texture and high-frequency information lack, which hinders convolutional neural network representation ability in image super-resolution. In order to solve this problem, a depth residual polarization image super-resolution network combined with double attention mechanism is proposed. This paper proposes a dual attention residual network model. The network structure is composed of a residual network with a global jump connection, which realizes the connection between the bottom network and the top network to stabilize the training of the deep network. The residual network contains 10 residual groups, and each residual group contains 20 residual blocks cascaded by dual attention blocks with local skip connections; At the same time, considering the interdependence between channels, an adaptive channel feature adjustment mechanism is designed. The channel attention mechanism is regarded as a guide, which biases the allocation of available processing resources towards the most informative part of the input; Cascaded spatial attention blocks are introduced to focus residuals characteristics on the key spatial contents. Spatial attention function measures the correlation between target features and key features, obtains the attention weight, and then aggregates the key content adaptively. The up-sampling module at the end of the network uses sub-pixel convolutional layers to reconstruct high-resolution images. The experiments mainly consist of two parts: first, use the bicubic degradation model to down-sample the training data set collected by the polarization camera, then add noise and blur to obtain the corresponding low-resolution image, and complete the training of the network model. The proposed method is compared with Bicubic, SRCNN, FSRCNN, EDSR methods to verify the effectiveness of the algorithm. In addition, the image reconstructed by this method is compared with the imaging system to provide data reference for system calibration and correction. The experimental results show that the reconstructed image with this method has richer texture details and uniform brightness, which is closer to the high-definition image of the imaging system. The peak signal-to-noise ratio and structural similarity index of this method are better than other methods, but the amount of parameters is only about 2/5 of that of EDSR. Thus, this method has the advantages of low information redundancy and better reconstruction effect.
