Joint Dense 3D Reconstruction Method for Endoscopic Images of Weak Texture Scenes

Abstract

Endoscopic inspection is an important non-destructive testing method. Traditional 3D endoscopic reconstruction methods, such as polarization reconstruction and shading reconstruction, have the drawbacks of not determining the object’s actual size and positional information. The stereo vision method is limited by its operating principles and has the issue of sparse reconstructed point clouds. These drawbacks greatly restrict the applications of the endoscope. Therefore, this work proposes a joint dense 3D reconstruction method for endoscopic images of weak texture scenes. This method uses the shading reconstruction normal to correct the polarization reconstruction normal, then uses coordinate conversion and point cloud fusion to convert the polarization and shading 3D reconstruction results from the pixel coordinate system to the world coordinate system. Finally combines the reconstruction results from the world coordinate system’s polarization, shading, and stereo vision. The fusion coefficients are obtained by solving the minimum error model, and then a complete and detailed 3D reconstruction surface was obtained in the world coordinate system. This method could avoid the difficulty of obtaining real coordinates for the 3D reconstruction of polarization and shading and prevent the issue of the sparse point cloud afforded by stereo vision reconstruction for weak texture scenes. The combined dense 3D reconstruction method had an average error of < 1% for length measurement of a 3D curve, which is highly significance for industrial endoscopic inspection.

Keywords

• Endoscopic detection
• dense point cloud
• 3D reconstruction