Development of optimal LWIR-to-color RGB transformation mapping based on neural nets using thermal IR hyperspectral imagery of scenes

OKSI will use hyperspectral data cubes collected at 128 bands within the thermal infrared spectral range of 7 to 14 microns to determine the optimal combination of three or four 1 µm wide bands that produce color (RGB) imagery that is most similar to that seen by the human eye. The resulting RGB images will be created from the chosen long wave infrared bands via a nonlinear mapping transformation using a biologically inspired neural network. The transformation map will be established on the basis of existing ground truth data (i.e., conventional color imagery and/or visible spectrum hyperspectral imagery) used in conjunction with corresponding thermal infrared hyperspectral imagery. OKSI already possesses a large archive of such hyperspectral thermal IR imagery, and has access to additional data from almost every hyperspectral LWIR imager in use in the US. Additional imagery will be generated, under a subsequent phase, using a high fidelity visible to LWIR scene generator such as DIRSIG. For real-time data processing, OKSI proposes to implement the mapping transformation in an FPGA. This provides a very low power, small package that can be built as an embedded hardware and combined with existing camera electronics. This approach benefits from other OKSI legacy projects in real time video data exploitation and color night vision development.