Shadows caused by non-uniform illumination can introduce significant errors in digital image correlation (DIC) measurements. The deep learning methods offer a good solution for image restoration. However, a dedicated deep learning approach for speckle images restoration in DIC measurement still requires further investigation. For this purpose, based on the quality assessment criterion of speckle images, this study aims to develop an improved Generative Adversarial Network (GAN)-based model for restoration of speckle images to eliminate the interference of non-uniform illumination on DIC measurement. Technically, the model consists of a generator with skip connections and multi-resolution discriminators to balance both global structures and fine details. In context of speckle images restoration, a speckle quality assessment module derived from the Gray-Level Co-occurrence Matrix (GLCM) is incorporated into the training, thereby guiding the generator to synthesize speckle patterns that more faithfully approximate real images in terms of brightness, contrast, and detail fidelity. In addition, a dataset is constructed with speckle images degraded by various non-uniform illumination and corresponding clear references. The Speckle-GAN model is evaluated with conventional image restoration methods. Both qualitative and quantitative evaluations demonstrate its superior capability in restoring clarity of speckle images degraded by non-uniform illumination shadows. Furthermore, DIC experiments under artificial non-uniform lighting are conducted, where the measurement errors are significantly reduced by applying the proposed method. The proposed method effectively mitigates illumination-induced degradation in speckle images, thereby improving accuracy and robustness of DIC under challenging lighting conditions.
Zhang et al. (Sat,) studied this question.