Enhance stereo-DIC in low quality speckle pattern by cross-scale stereo matching with application in dental crack detection

• Proposed a cross-scale semi-global stereo matching algorithm with geometric constraints (CS-GCSGM) which combines geometric constraint-based semi-global stereo matching (GC-SGM) with cross-scale cost aggregation (CSCA), effectively improving the accuracy of the initial guess under low-quality speckle patterns . • Validated algorithm accuracy and robustness on speckle patterns with different mean intensity gradients (MIG), demonstrating that CS-GCSGM achieves higher accuracy and robustness on low-quality speckle patterns through comparison on representative algorithms within the path-independent category (FFTCC, SIFT and GC-SGM). • The proposed CS-GCSGM showcases the extensive potential of the method in practical application such as tooth crack recognition. Stereo-digital image correlation (DIC) is a full-field deformation measurement technique. The convergence of current iterative DIC relies on an accurate initial guess, and the precision and reliability of the initial guess depend on the quality of the speckle pattern. However, due to the limitations in speckle pattern fabrication techniques and the experience of operators, it is common to encounter low-quality speckle pattern. Dealing with low-quality speckles poses a significant challenge in the field of DIC. In this study, a cross scale semi-global stereo matching algorithm with geometric constraints (CS-GCSGM) is proposed, which combines geometric-constraint-based semi-global stereo matching (GC-SGM) with cross-scale cost aggregation (CSCA). This algorithm aims to enhance the accurate of the initial guess for the inverse compositional Gauss-Newton (IC-GN) algorithm on low-quality speckles pattern. To validate the effectiveness of the proposed algorithm, speckle patterns with various mean intensity gradient (MIG) were fabricated on both simple shapes and complex surface objects. A set of stereo matching and 3D reconstruction experiments were conducted on these speckle patterns to assess the accuracy of the proposed method. Experiment results reveal that the CS-GCSGM algorithm achieves higher accuracy and robustness with low-quality speckle patterns than other algorithms, while also demonstrating optimization in high-quality speckle patterns. In the final experiment of tooth crack recognition, the successful integration of 2D-DIC and CS-GCSGM effectively achieved 3D reconstruction. This experiment prominently showcases the extensive application potential of the proposed method.