One approach would be to use 3D reconstruction and then use deformable ICP (iterative closest point) to recover the deformation map of the embossed text. The deformation map would transform the CAD model to the deformed scene. The inliers you compute during correspondence estimation would also allow you to segment these characters. (This might be similar to computing optical flow)
For reconstructing your scene, I would recommend photometric stereo techniques. They are pretty accurate in reconstructing small details. For large objects, they would not work. Some intro texts are here and here.
Also, in such cases, you could use an appropriate lighting to emphasize the text. A low angle ring light would be an example.