Super-resolution of photoelastic polarization images using a neural network and a divide-and-conquer strategy

Abstract

The use of polarization cameras in digital photoelasticity has become increasingly prevalent, significantly contributing to the advancement of digital photoelasticity. However, polarization images captured by polarization cameras suffer from resolution loss, which severely affects the accuracy of the isochromatics and isoclinics. Therefore, it is essential to enhance the resolution of these polarization images. In this study, a super-resolution reconstruction network specifically designed for photoelastic polarization images is proposed for the first time. A large dataset was generated and collected to support the development of this network. Additionally, a divide-and-conquer strategy is introduced to avoid interference between the different degradation models. This strategy also enables the efficient acquisition of tens of thousands of real-world datasets. Comparative experiments demonstrate that the proposed divide-and-conquer strategy significantly enhances the network's accuracy. In comparison with interpolation methods and existing networks, the proposed network consistently outperforms them, showing superior performance. This network effectively increases the resolution of photoelastic polarization images, leading to more accurate isochromatics and isoclinics.