An evaluation of convergence criteria for digital image correlation using inverse compositional Gauss-Newton algorithm

A fast, robust and accurate digital image correlation (DIC) method, which uses a robust zero-mean normalized sum of squared difference correlation criterion, a sophisticated reliability-guided displacement tracking strategy and an efficient inverse compositional Gauss-Newton (IC-GN) algorithm, was recently proposed for full-field deformation measurement. As an iterative local optimization algorithm, IC-GN algorithm iteratively solves for the incremental warp assumed on the reference subset until the preset convergence criteria are satisfied. In the literature, different convergence criteria have been set for iterative optimization algorithms. However, on the one hand, stringent convergence criteria lead to increased number of iterations and lessen the computational efficiency. On the other hand, too loose convergence conditions enhance the computational efficiency but may decrease the registration accuracy. Understanding the impact of prescribed convergence criteria on DIC measurement and how to choose proper convergence criteria are therefore fundamental problems in realizing high-efficiency yet high-accuracy DIC analysis. In this paper, the convergence characteristics of IC-GN algorithm are investigated in terms of convergence speed and radius of convergence using real experimental images. The effect of various convergence criteria on the efficiency and accuracy of IC-GN algorithm are carefully examined. Recommendations are given to select proper convergence criteria for more efficient implement of IC-GN algorithm.