Single-image-based nonuniformity correction of uncooled long-wave infrared detectors: a deep-learning approach

Area of Science:

• Infrared imaging technology
• Optoelectronics
• Machine learning applications

Background:

• Fixed-pattern noise (FPN) is a significant challenge in infrared imaging systems.
• FPN arises from non-uniform responses in microbolometer focal-plane-array (FPA) optoelectronics.
• Current non-uniformity correction (NUC) methods often require handcrafted operators.

Purpose of the Study:

• To develop an improved single-image-based non-uniformity correction (NUC) operator for infrared imaging.
• To demonstrate that NUC operators can be learned directly from simulated training data using deep learning.
• To reconstruct noise-free infrared images from noisy observations.

Main Methods:

• Utilized a comprehensive column FPN model to simulate nonlinear characteristics of FPA readout circuits.
• Generated a large dataset of high-fidelity simulated infrared images.
• Employed an end-to-end residual deep network (CNN) for learning FPN characteristics.
• Estimated and subtracted column FPN from raw infrared images.

Main Results:

• The proposed deep-learning-based approach effectively reconstructs noise-free infrared images.
• Accurate estimation and subtraction of column FPN were achieved.
• Demonstrated superior performance in FPN removal compared to state-of-the-art methods.
• Showcased enhanced detail preservation and artifact suppression in real-world infrared images.

Conclusions:

• A novel, data-driven NUC operator can be learned using CNNs and FPN simulation.
• This deep learning technique offers a powerful alternative to handcrafted NUC methods.
• The approach significantly improves the quality of infrared images by effectively removing FPN.