Background <p>Stereo Digital Image Correlation (Stereo-DIC) is a non-contact optical three-dimensional measurement technique based on speckle image matching. It enables three-dimensional shape and deformation measurements through inter-frame and inter-camera matching. In recent years, the rapid development of deep learning has had a significant impact on the field of image matching, and there has been increasing attention on using neural networks to enhance the efficiency and accuracy of DIC measurements. However, existing methods face two challenges: First, there is no unified Stereo-DIC network to achieve temporal and stereo matching simultaneously. Second, the black box characteristic of neural networks leads to image matching lacking interpretability, which hinders the wide application and further research of these methods.</p> Objective <p>To solve these problems, this paper proposes a unified and physics-guided speckle matching network for Stereo-DIC, referred to as Stereo-DICNet2.</p> Methods <p>Stereo-DICNet2 only uses deep learning to extract image features and proposes a feature-based cross-correlation (FCC) matching layer that performs matching through specific mathematical operations rather than a black box. In addition, a lightweight RAFT (Recurrent All-Pairs Field Transforms) iterative optimization algorithm is introduced to optimize the displacement.</p> Results <p>Stereo-DICNet2 can achieve micron-level three-dimensional measurement accuracy in material tensile experiments.</p> Conclusions <p>The proposed method explicitly integrates deep learning with mathematical operations to address the dual challenges of unified spatiotemporal matching and interpretability limitations, demonstrating high practical value in experimental validations.</p>