Single-image-based Nonuniformity Correction of Uncooled Long-Wave Infrared Detectors: a Deep-Learning Approach.
Zewei He,Yanpeng Cao,Yafei Dong,Jiangxin Yang,Yanlong Cao,Christel-Loic Tisse
DOI: https://doi.org/10.1364/ao.57.00d155
IF: 1.9
2018-01-01
Applied Optics
Abstract:Fixed-pattern noise (FPN), which is caused by the nonuniform opto-electronic responses of microbolometer focal-plane-array (FPA) optoelectronics, imposes a challenging problem in infrared imaging systems. In this paper, we successfully demonstrate that a better single-image-based non-uniformity correction (NUC) operator can be directly learned from a large number of simulated training images instead of being handcrafted as before. Our proposed training scheme, which is based on convolutional neural networks (CNNs) and a column FPN simulation module, gives rise to a powerful technique to reconstruct the noise-free infrared image from its corresponding noisy observation. Specifically, a comprehensive column FPN model is utilized to depict the nonlinear characteristics of column amplifiers in the readout circuit of FPA. A large number of high-fidelity training images are simulated based on this model and the end-to-end residual deep network is capable of learning the intrinsic difference between undesirable FPN and original image details. Therefore, column FPN can be accurately estimated and further subtracted from the raw infrared images to obtain NUC results. Comparative results with state-of-the-art single-image-based NUC methods, using real-captured noisy infrared images, demonstrate that our proposed deep-learning-based approach delivers better performances of FPN removal, detail preservation, and artifact suppression.