This goal is to design a novel framework for thermal image restoration based on camera modeling. The physical imaging process is first analyzed resulting in modeling the thermal camera system as a combination of linear spatial convolution with point spread function (PSF) and scalar distortion by a nonlinear function. PSF is then reconstructed to be a Gaussian function approximating the image of a point source of heat, while the nonlinear function is estimated to be the one whose inverse minimizes the error in conforming to the heat diffusion regularity. Hence, linearized by the inverse of the nonlinear function and then filtered by the inverse of PSF, the thermal image can be restored.

PSF image acquisition: An ice cube covered by a heated CDROM is placed far away from the thermal camera, so that the ice "in the ROM hole" is mapped to be a black pixel in the thermal image. It is then turned to be an isotopic point source.  PSF approximation: covariance matrix, rather than a unique variance value, is used as the unknown parameter in the approximating Gaussian model to account for the situation when PSF is not isotopic.

Three regularized inverse techniques: Tikhonov, Total Variation (TV) and Lasso are used to apply the PSF to restore the thermal image. TV and Lasso excel in being capable of preserving the edge features without amplifying the noise. Compared with TV, Lasso is executed much faster.