Image matching, which is a cornerstone in 3D reconstruction, poses significant and considerable challenges when applied to diverse, unstructured image collections. This paper presents a robust methodology employing SuperGlue and DKM to address these challenges. Initially, EfficientNet is employed to retrieve matching image pairs based on global feature extraction. Subsequently, SuperGlue identifies correspondences between 2D features across these image pairs, while DKM manages large geometric deformations and appearance variations, enhancing the robustness and precision of the matching process. In the Kaggle competition Image Matching Challenge 2023, our approach demonstrated its effectiveness, securing the 34th position among 494 teams and achieving a mean Average Accuracy (mAA) score of 0.470. This achievement not only underscores the potential of our method in real-world applications, particularly in 3D reconstruction from unstructured image collections, but also contributes to the advancement of efficient 3D reconstruction techniques. The implications of this work extend to various applications, including mapping services, cultural heritage preservation, and numerous online services, paving the way for future research in 3D vision tasks and emphasizing the importance of robust image matching techniques in the broader field of computer vision.