Feature Fusion-Enhancement: A Novel Solution for 2D-3D Reconstruction in Feature-Uniform Contexts

Vision-based 3D reconstruction technology is a method that uses 2D images to restore the 3D structural information of objects. It has extremely important research value and application prospects in the field of medical image processing. However, due to the lack of obvious texture and high-contrast edges in tooth images, and the single color and similar geometric shapes of tooth surfaces, the existing 3D reconstruction methods perform poorly in tooth scenes. Therefore, this paper proposes the EESD-H (Edge-Enhanced SIFT with Hue Descriptor) algorithm, which integrates the edge features and color features of tooth images to improve the accuracy of feature point extraction and matching in the algorithm to deal with the problems of poor texture information and single color in tooth scenes; at the same time, a weighted EPnP-WGN algorithm is proposed, which improves the accuracy of pose estimation in tooth scenes by introducing object-space residuals as weights in the EPnP solution process; in addition, a PatchMatch Stereo algorithm based on bilateral weight function as matching cost is proposed, which improves the accuracy of depth information estimation of tooth image sequences by combining spatial proximity function and color similarity function in HSV space as the matching cost of PatchMatch Stereo algorithm. Experimental results show that our algorithm has excellent results in the Three-Dimensional Reconstruction of Dental Structures from RGB Image Sequences task. The F-score obtained after overall algorithm optimization is improved by more than 30% compared with that before optimization.