Improving Resolution of Translated Infrared Images

Michiya Kibe, Takeru Inoue, Junya Morioka, Ryusuke Miyamoto

Abstract: With the development of generative adversarial networks (GANs), it has become possible to perform style transfers from infrared images, which often lack color and contrast information, to more familiar domain images. For sensor systems, it is crucial to ensure the performance of transference to the target domain at the resolution required to recognize detailed structures. In this study, we propose the application of a novel structural loss function that excludes the luminance component from the structural similarity index measure (SSIM) to enhance style translation accuracy and resolution. To achieve this, we conducted experiments by translating mid-wave infrared (MWIR) images to shortwave infrared (SWIR) images, which are easier for humans to interpret. The experimental results show that the introduction of cycle consistency and SSIM loss is effective for improving the performance of the modulation transfer function (MTF) and that the application of the proposed structural loss function further enhances performance.