Low-Frequency Nonuniformity Correction in Static Thermal Images

A frequent issue in uncooled thermal cameras is the presence of a low-frequency shading or non-uniformity (NU), where slowly spatially varying changes in intensity corrupt the radiometric image. Usual correction methods for this problem rely on motion in the scene and are therefore unsuitable for static cameras and for restoring individual images. Depending on its physical origin, the NU can be multiplicative, additive, or in-between these two extremes. We propose a static-image demixing method where we separate the low frequency component causing the NU from the underlying “true” image. Our contribution is three-fold: 1) we propose a parametric transformation that allows a subtractive demixing regardless of the multiplicative or additive nature of the NU; 2) we design a cost functional to evaluate candidate estimates of the NU and of the multiplicative/additive mixing parameter; 3) we propose an iterative method where the NU estimate is progressively updated by optimizing a parametric perturbation with respect to the cost functional. In spite of its simplicity, our method results in a nonparametric NU estimate and a nonlinear demixing. Experiments on simulated and real thermal imagery demonstrates that it successfully removes the low frequency shading from static scenes. Individual iterations can be also interleaved between frames of a video, allowing for continuous adaptation to changes in the NU.