Natural wooden cream sticks exhibit diverse depth defects on their surfaces, presenting significant challenges for automatic detection. Current inspection technique mainly relies on manual means, which cannot provide quantitative evaluation and significantly limits production efficiency. Realizing in situ, online, and intelligent detection of surface depth defects in ice cream sticks is crucial for enhancing automation and ensuring consistent product quality. Through considering the surface and geometric features of ice cream sticks, the paper proposes a binocular three-dimensional (3D) reconstruction method for ice cream sticks, which incorporates speckle projection and Segment Anything Model v2(SAM2)-based target mask extraction. By performing automatic mask generation, candidate region selection, and refined mask extraction, the region of interest (ROI) of target is obtained from binocular speckle images. A colour-gradient weighted stereo matching algorithm combined with guided filtering generates high-quality disparity maps, and achieves a matching efficiency improvement exceeding 45 times. The 3D reconstruction experiment of a textureless step-shaped standard block verifies that the proposed method can achieve the measurement accuracy of about 0.05 mm. For various defective ice cream sticks, the reconstructed results compared with manual measurements show a maximum error not exceeding 3.4%. The proposed method provides an efficient and reliable 3D reconstruction solution for the in situ, non-destructive detection of depth defects in ice cream sticks, demonstrating practical value and promising engineering applicability.
To address the challenges of unreliable feature matching, high mismatch rates, and limited reconstruction accuracy in binocular stereo vision applied to asphalt pavement with inherent weak texture features, this paper proposes a speckle-assisted binocular 3D reconstruction method based on a multi-scale adaptive feature fusion algorithm. Infrared speckle patterns are actively projected to enrich the pavement surface features, and a multi-scale matching framework is developed by integrating Laplacian pyramid representations, feature-driven adaptive regularization, and Softmax-based nonlinear fusion. This design can achieve stable and accurate disparity estimation, even in weak texture regions, and produce high-quality 3D point clouds that faithfully represent both macro-scale undulations and micro-scale texture details. Ablation experiments validate the effectiveness of the proposed modules, showing that the relative errors of the arithmetic mean height (Sa) and root-mean-square height (Sq) are reduced to below 2.3%. When aligned with 3D scanner data using the iterative closest point (ICP) algorithm, the reconstructed point clouds achieve sub-millimeter mean error and an overlap rate exceeding 97%. The results indicate that the proposed method offers a reliable technical solution for efficient texture-depth analysis and practical pavement condition assessment.
To address the challenges of unreliable feature matching, high mismatch rates, and limited reconstruction accuracy in binocular stereo vision applied to asphalt pavement with inherent weak texture features, this paper proposes a speckle-assisted binocular 3D reconstruction method based on a multi-scale adaptive feature fusion algorithm. Infrared speckle patterns are actively projected to enrich the pavement surface features, and a multi-scale matching framework is developed by integrating Laplacian pyramid representations, feature-driven adaptive regularization, and Softmax-based nonlinear fusion. This design can achieve stable and accurate disparity estimation, even in weak texture regions, and produce high-quality 3D point clouds that faithfully represent both macro-scale undulations and micro-scale texture details. Ablation experiments validate the effectiveness of the proposed modules, showing that the relative errors of the arithmetic mean height (Sa) and root-mean-square height (Sq) are reduced to below 2.3%. When aligned with 3D scanner data using the iterative closest point (ICP) algorithm, the reconstructed point clouds achieve sub-millimeter mean error and an overlap rate exceeding 97%. The results indicate that the proposed method offers a reliable technical solution for efficient texture-depth analysis and practical pavement condition assessment.