Objective In view of the limitations in quantity and quality when endoscopes acquire internal human body images this study proposes a new high-definition endoscope video frame extraction strategy for 3D reconstruction. Methods First a composite method based on the Structural Similarity Index Measure SSIM and the Laplace operator was used to screen and extract key frames with high definition and low noise. Then the Neighbor2Neighbor deep-learning method was used to process the preliminary key-frame queue to reduce noise in the images and enhance their quality. The COLMAP incremental Structure-from-Motion SFM technique was employed to convert high-quality key frames into sparse point-cloud data. Finally the OpenMVS depth-map fusion technology was used for dense reconstruction and surface fitting to obtain a high-precision three-dimensional model. Results The experimental results showed that the proposed method not only significantly enhanced the image feature matching degree but also greatly improved the accuracy and precision of three-dimensional reconstruction. Conclusion This method not only significantly improves the accuracy of image feature matching but also enhances the accuracy and precision of three-dimensional reconstruction. It is expected to provide clinicians with a more intuitive diagnostic basis.