DIC-Net: Upgrade the performance of traditional DIC with Hermite dataset and convolution neural network

Abstract

Digital image correlation (DIC) is a non-contact optical method that tracks the speckle pattern on specimen surface to calculate the displacement and strain by image correlation algorithm. Although the traditional DIC method can conveniently measure surface deformation, it still has many limitations: (1) the accuracy of displacement and strain calculation needs to be improved in the case of high deformation gradient; (2) under match or over-match can hardly be avoided when the filters are used to reconstruct smooth displacement or strain field, and (3) boundary effect remains unresolved in computing the deformation near the boundary of region of interest or the discontinuous area (e.g. area near crack tip or crack face). Recently, the deep learning based DIC (Deep-DIC) has revealed its attractive ability in handling above issues in traditional DIC, and impressive results have been achieved. The mean value of the absolute error (MAE) on the test set has been optimized to 0.0361 pixels using existing Deep-DIC approaches, which are accompanied by a real-time measurement speed. The network structure and training dataset are two key factors for the deep learning method. However, the current working networks have been modified from other image tasks and cannot fully adapt to the demands of the DIC tasks, and the dataset they generated still has evident flaws, limiting the method's accuracy and generalization performance which is utilized to assess performance on samples outside the training set. In this paper, we firstly proposed a new Hermite dataset that is created by using the high-order Hermite element to take account more complex deformation, then a new network architecture designed for the DIC task has been developed to extract richer deformation features. A test set of 3216 examples containing six different modes of displacement is used to compare the performance of our network with others. The proposed DIC-Net-d achieves the lowest MAE in the test set. Meanwhile, in the Star5 image sets from DIC-Challenge, the proposed DIC-Net-d achieves a spatial resolution of 17.25 pixels and a noise level of 0.0136 which outperforms existing traditional and non-traditional methods. Finally, the strain network trained by our Hermite dataset is also successful in predicting the strain field of Star6 in the DIC challenge. The experiment results show the superiority of the proposed Hermite dataset and new network with respect to other Deep-DIC methods.