Abstract: 3D reconstruction is the essential technique of retrieving the three dimensional structure from a two-dimensional image. The process entails finding distinctive spots within the image and subsequently matching them to build positional correlations among other photos. To recreate the three dimensional structure from two-dimensional pictures, this criterion must be fulfilled. The efficacy of feature identification and matching algorithms directly affects the quality and comprehensiveness of 3D reconstruction. This essay first analyzes the traditional algorithms SIFT and ORB and offers a brief summary of the challenges linked to these established methods. You should compare the feature point detection and matching algorithms SuperPoint and SuperGlue within the existing deep learning framework. An effective framework for feature point identification and matching was established by integrating two enhanced networks. Numerous experiments were performed on the Hmatches dataset, revealing that the enhanced network surpassed the traditional SIFT algorithm, increasing the fitting accuracy of the transformation matrix by 5.168% and 3.767%, the correctly matched point pairs by 40.18%, and the matching accuracy by 10.16%.