抄録
This paper proposes a self-calibration and neural network based photometric stereo for virtual image generation,
and purposes an improvement in precision of generating image.
This method can generate virtual images without an calibaration object which has the same reflectance function as the test object under the same light source direction. However, the accuracy for generated virtual image for improvement requires a large number of images in the previous neural netwok based rendering.
The proposed method uses the angle between a unit surface
normal vector and an equally divided vector between each light source direction and viewing direction, and the approach can increase efficiency for neural network learning. Better virtual image can be generated from a few observed images. Generated images are evaluated using Phong model and Torrance-Sparrow model in comparison with the previous approach.
