A structure prior weighted hybrid l(2)-l(p) variational model for single infrared image intensity nonuniformity correction

Infrared (IR) image nonuniformity correction is of critical importance to visual perception and subsequent applications. A range of methods have been extensively studied, whereas they either fail to achieve visually natural results, or only exhibit high performance for a specified type of nonuniform scene (e.g., Gaussian-shaped aero-thermal radiation-induced intensity nonuniformity). In the present study, a novel variational framework is proposed, reasonably considering the inherent characteristics both of the latent IR image and bias field, as an attempt to correct IR image intensity nonuniformity. In other words, given the gradient correlation between the degraded image and the bias field, a structure prior weighted l(2)-norm regularizer is substituted to constrain the bias field. Moreover, under the global sparseness of latent clear IR image, a l(p)-regularized term is employed to accurately restore the latent image, as well as eliminating texture noise. To efficiently optimize this hybrid l(2)-l(p) variational model, both iteratively re-weighted least square and alternating minimization schemes are adopted to address the mentioned non-convex problem. According to the experiments of several simulated and real IR datasets, the feasibility of the proposed algorithm and its ability to fit different nonuniform conditions compared with five existing methods are demonstrated.