Deep learning-based digital volume correlation

Abstract

With the rapid development of 3D volume imaging technology, digital volume correlation (DVC) has already made great progresses and become an indispensable tool in various engineering kingdoms. Inspired by artificial intelligence-related technologies, we develop a novel deep learning-based DVC methodology for 3D deformation analyses with the aid of a set of well-trained neural networks, hereafter referred to as DVC-Net, to implement accurate, precise and robust 3D deformation characterization in an efficient manner in this work. This proposed DVC-Net comprises three sub-networks, i.e., DVC-NetI, DVC-NetS and DVC-NetD, which are used in turn for integer-voxel searching, sub-voxel registration and displacement field denoising. Unlike traditional DVC methods, e.g., subvolume-based local DVC or finite-element method-based global DVC, the proposed DVC-Net neither explicitly relies on correlation criteria nor requires numerical iterative calculations, which greatly reduces computational complexity of the DVC analyses and therefore improves computational efficiency by at least 2∼3 orders of magnitude. Besides, it is suitable for dealing with various practical problems arising in the DVC-related studies such as volumetric images that experience changes in lighting conditions or noisy images. It is anticipated that the proposed DVC-Net can provide an alternative method for high-performance volumetric deformation characterization.