Registration of infrared transmission images using squared-loss mutual information

• We address the problem of registering infrared transmission images.
• We propose squared-loss mutual information as a similarity measure for images.
• The proposed method is applied to reconstruct the inner structure of semiconductor.
• Experiments show the practical usefulness of the proposed method.

Infrared light allows us to measure the inner structure of opaque samples such as a semi-conductor. In this paper, we propose a method of registering multiple infrared transmission images obtained from different layers of a sample for 3D reconstruction. Since an infrared transmission image obtained from one layer is contaminated with defocused images coming from other layers, registration with a standard similarity metric such as the squared error and the cross correlation does not perform well. To cope with this problem, we propose to use the squared-loss mutual information as an alternative similarity measure for registration, which is more robust against noise than ordinary mutual information. The practical usefulness of the proposed method is demonstrated in simulated and actual experiments.