Convolutional Neural Network Based Denoising for Digital Image Correlation Reconstructing High-Fidelity Deformation Field

Digital image correlation (DIC) is a widely used technique for full-field deformation measurement based on image processing. The high quality of images input for correlation is the most important guarantee for obtaining accurate physical field. In this paper, we propose a new method to deal with the negative effects of noise especially for deep learning-based DIC. Denoising convolutional neural network (DnCNN) block is introduced to deal with the noise before speckle patterns are fed into deep learning-based DIC. A new speckle pattern dataset, SPDataset, is created to train DnCNN on speckle features rather than on images with semantic information. The performance of networks trained on specific and blind noise level is compared and the possibility of network structure simplification is explored. The depth and performance of the network reach a good compromise at a layer number of 9. The denoising ability of the network for speckle patterns was confirmed qualitatively and quantitatively.