Three-dimensional reconstruction techniques are used in various situations such as visual inspections in a factory and digital archiving in a museum. One such method is photometric stereo, in which a change in the direction of the light source is used to illuminate an object. This method can reconstruct objects with high accuracy. By contrast, this approaches have certain limitations, such as the requirement for a dark- room environment with no more light sources than necessary, and the need for a Lambert diffuse reflection material that reflects the incident light equally in all reflection directions. Therefore, a method for transforming the specular reflection components of an object into a diffuse form using deep learning was previously developed. However, this method applies a network in which the object surface is not changed de- spite the change in the position of the light source. It is therefore not applicable to a shape reconstruction using photometric stereo, which ap- plies a change of the position of the light source. In this study, we propose an improved learning network that transforms captured specular images into a diffuse form while reflecting the changes in the surface luminance by applying images that differ in the position of the light source and its corresponding data. In addition, we verified the accuracy and effectiveness of a conventional method for both the transformed 2D images and the 3D shape model reconstructed from these images. As a result, we confirmed that the accuracy of our method is improved in both cases.
View full abstract
