Digital image correlation (DIC) has become an industry standard to retrieve accurate displacement and strain measurement in tensile testing and other material characterization. Though traditional DIC offers a high precision estimation of deformation for general tensile testing cases, the prediction becomes unstable at large deformation or when the speckle patterns start to tear. In addition, traditional DIC requires a long computation time and often produces a low spatial resolution output affected by filtering and speckle pattern quality. To address these challenges, we propose a new deep learning-based DIC approach -- Deep DIC, in which two convolutional neural networks, DisplacementNet and StrainNet, are designed to work together for end-to-end prediction of displacements and strains. DisplacementNet predicts the displacement field and adaptively tracks the change of a region of interest. StrainNet predicts the strain field directly from the image input without relying on the displacement prediction, which significantly improves the strain prediction accuracy. A new dataset generation method is proposed to synthesize a realistic and comprehensive dataset including artificial speckle patterns, randomly generated displacement and strain fields, and deformed images based on the given deformation. Proposed Deep DIC is trained purely on a synthetic dataset, but designed to perform both on simulated and experimental data. Its performance is systematically evaluated and compared with commercial DIC software. Deep DIC gives highly consistent and comparable predictions of displacement and strain with those obtained from commercial DIC software, while it outperforms commercial software with very robust strain prediction even with large and localized deformation and varied pattern qualities.