Noise-Robust Image Deblurring Using PSF Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image deblurring methods are inadequate for noise-robust image deblurring, particularly in long wavelength infrared (LWIR) images, as they rely on bright point-like sources that are impractical for uncooled LWIR imaging systems and introduce additional image blur when using median filtering.
Innovation Solution
A method that iteratively denoises reference and scene images while preserving straight edges, estimates a point spread function (PSF) using noise-regularized inversion of an integration operator, and deblurs images using the estimated PSF, avoiding median filtering and derivative-based edge detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If bright point-like sources are used to measure PSF, then PSF measurement accuracy is improved, but the method becomes impractical for uncooled LWIR imaging systems and introduces additional image blur
Solution Approach 1:
The patent extracts the essential information needed for PSF measurement (edge transition characteristics) from a practical scene element (straight edge) rather than requiring an idealized point source. This allows PSF estimation using naturally occurring edges in the scene that are captured by the imaging system itself, eliminating the need for special point source targets.
Solution Approach 2:
The patent uses the imaging system's own captured image of a straight edge to create a measurement reference, rather than requiring an external point source. The edge information is copied from the actual imaging scenario, making the measurement process consistent with real operating conditions and avoiding additional blur from point source requirements.
2Object-affected harmful factors
If median filtering is applied to reduce noise, then noise reduction is improved, but additional image blur is introduced
Solution Approach 1:
The patent applies different processing qualities to different regions: noise reduction is applied in regions away from edges, while edge regions preserve their sharpness through specialized handling. The edge spread function extraction focuses on the transition region, applying minimal processing that maintains edge integrity while reducing noise in surrounding areas.
Solution Approach 2:
The patent segments the image processing into distinct regions: edge detection identifies the transition zone, the edge spread function is extracted specifically from this zone, and noise reduction is applied to non-edge regions. This segmentation allows noise reduction without compromising edge sharpness, as each region receives appropriate processing.
3Difficulty of detecting and measuring
If derivative-based edge detection is used, then edge detection capability is improved, but noise sensitivity increases
Solution Approach 1:
Instead of detecting edges by finding where derivatives are maximum (traditional approach), the patent inverts the approach by integrating the image intensity to create an edge spread function, then differentiating this integrated function. This integration-first approach naturally suppresses noise before differentiation, reducing noise sensitivity while maintaining edge detection capability.
Data Source
AI summary
Systems and methods related to pre-processing images and estimating PSFs for image deblurring in a noise-robust manner. The pre-processing may include iteratively denoising a reference image, iteratively denoising a scene image containing at least one straight edge within the scene image while preserving the at least one straight edge. The pre-processing may include smoothing, subtracting and dividing to provide a pre-processed image. The PSF-estimation process may include generating, with the at least one processor, a first noise value from portions of an edge intensity profile, generating a second noise value from a difference between the edge intensity profile and an integral of the noise-regularized inversion of the integration operator operating on the edge intensity profile and determining a value of a local curvature regularization parameter resulting in the first noise value and the second noise value being within a tolerance range.