Deep learning based speckle image super-resolution for digital image correlation measurement

Abstract

Digital image correlation (DIC) is a non-contact deformation measurement method based on speckle matching, widely used in experimental mechanics, explosive mechanics, construction measurement and other fields. However, when the DIC method uses a small resolution camera to measure large-sized objects, the resolution of speckle images will decrease. This not only leads to a decrease in the resolution of the measured deformation field, but also reduces the speckle size in the image, resulting in a decrease in measurement accuracy. To improve the resolution of the speckle image, we propose a deep learning-based speckle image super-resolution approach, named Speckle-SRGAN. Speckle-SRGAN is designed based on the high-frequency and fine texture characteristics of speckle images, and it introduces coordinate attention mechanism and global depth residual module to preserve high-frequency and fine textures. Low resolution speckle images are processed by Speckle-SRGAN to transform into high-resolution speckle images with high fidelity. Simulation and experimental results show that Speckle-SRGAN can increase the resolution of speckle image by 4 times and the speckle is smooth without loss of details. The real experiment also shows that using our method to preprocess speckle images can reduce the measurement error of traditional DIC methods by about 0.01 pixels. The code and data of this paper is released at: SpeckleSRGAN.