Nonuniformity Correction Based on Adaptive Sparse Representation Using Joint Local and Global Constraints Based Learning Rate

The infrared focal plane array (IRFPA) suffers from the problem of non-uniform photoelectric responsivity of infrared detectors, namely nonuniformity. This paper aims at removing the fixed pattern noise (FPN) brought by nonuniformity in IRFPA through scene-based nonuniformity correction (SBNUC). Different from traditional neural network-based methods where spatial filters are adopted, a new approach of acquiring the desired image (the target of optimization task) is proposed, which is based on the sparse representation theory. The sparse representation of the image corrected by the current correction coefficients is used as the desired image. During sparse coding, the error tolerance is computed from an adaptive criterion, which is negatively correlated to the root-mean-square error between the input image and the corrected image. The resulted desired image is more accurate than its counterparts in traditional filter-based methods, as the edges can be preserved without heavily blurring image details, while the elements of FPN are omitted as redundant when coding the representation coefficients. To suit this desired image, a new adaptive learning rate based on joint local and global constraints is proposed. Different from traditional local spatial standard deviation-based learning rate, the constraints are calculated based on the variations among the input image, the corrected image, and the desired image. A new motion detection rule is also incorporated to this learning rate, which could be maximized to accelerate convergence only if the scene motion is sufficient both globally and locally. Experiments are conducted on both public infrared image sets (the Video Verification of Identity program data set and the Automonous Systems Lab data set) and a private infrared sequence. Experimental results demonstrate that, the proposed method achieves better performance of FPN removal on infrared sequences with both simulated and real nonuniformity, compared with SBNUC methods based on spatial filter (including Gaussian filter, guided filter, and bilateral filter).