Adaptive Multi-Interpolation and Multidimensional Feature Weighting for Binocular 3-D Visible Surface Reconstruction

Abstract

To address the insufficient matching accuracy of traditional stereo matching in areas with weak textures, low contrast, and edges of millimeter-scale miniature components, this article proposes a binocular 3-D reconstruction method for visible surfaces based on adaptive multi-interpolation and multidimensional feature weighting. First, the method integrates bicubic and Lanczos interpolation, generating adaptive super-resolution images to enhance the characterization of texture details. Second, in the cost computation stage, a novel adaptive cross-window is implemented, which combines the gray-level co-occurrence matrix (GLCM) and the Scharr operator for joint texture feature extraction. Simultaneously, neighborhood weights are dynamically adjusted based on color, spatial distance, and gradient information to improve matching robustness in depth-discontinuous regions. Finally, by analyzing the characteristics of the discrepancy between the reconstructed height and the actual height, an error compensation function is constructed to achieve high-precision compensation for the visible surface reconstruction. Experimental results demonstrate that the proposed method achieves strong performance on the Middlebury dataset, with average mismatch rates of 2.547% in nonoccluded regions and 5.041% in all regions, while also maintaining high accuracy across different measurement ranges: the average absolute error reaches 0.044 mm for measurements below 5 and 0.080 mm for the 5–10-mm range.