Frequency Domain Noise Removal for Light Field Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional spatial domain-based image denoising methods blur image details while attempting to remove periodic noise from reconstructed light field images, which are caused by stitching edges during the reconstruction process.
Innovation Solution
A frequency domain-based method that involves acquiring a light field image, calibrating microlens imaging centers, transforming the image into the frequency domain, applying a low-pass filter to suppress high-frequency periodic noise, and using an image mask to filter out noise components, thereby preserving image details.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If spatial domain-based denoising methods (mean filtering, median filtering, Gaussian filtering) are used to remove periodic noise from reconstructed light field images, then noise removal effect is improved, but image details are blurred and lost
Solution Approach 1:
The patent transforms the image denoising problem from the spatial domain to the frequency domain using Fourier transform. By changing the dimension of analysis from spatial coordinates to frequency components, the method can selectively remove periodic noise (which appears as specific frequency components) without affecting image details (which are distributed across multiple frequencies). This dimensional transformation resolves the contradiction by enabling targeted noise removal while preserving useful image information.
Solution Approach 2:
The patent applies different processing strategies to different frequency components in the frequency domain. Low-frequency components (which contain image structure and detail information) are preserved, while specific high-frequency components (which correspond to periodic noise from stitching edges) are removed. This selective, localized processing in the frequency domain allows noise removal without blurring image details, resolving the contradiction between noise removal and detail preservation.
2Ease of manufacture
If pixel blocks of the same size are taken from microlens images and stitched together for reconstruction, then reconstruction process is simplified, but stitching edge noise is generated with periodic arrangement
Solution Approach 1:
The patent extracts and removes the harmful periodic noise components from the reconstructed image by transforming to the frequency domain. The Fourier transform converts the spatial periodic noise into distinct frequency peaks, which are then identified and removed using spectral analysis. This extraction approach separates the noise from the useful image content, allowing simplified reconstruction processing to continue while eliminating the generated stitching edge noise.
3Loss of time
If microlens images are stitched directly without calibration to obtain reconstructed images, then processing time is reduced, but imaging center misalignment causes reconstruction accuracy degradation
Solution Approach 1:
The patent performs microlens imaging center calibration before the reconstruction process. By determining and storing the precise imaging centers of each microlens in advance, the calibration results can be reused for multiple reconstructions, reducing processing time for subsequent operations. This preliminary action ensures high reconstruction accuracy without significantly increasing overall processing time, as the calibration is performed once rather than for each reconstruction.
Data Source
AI summary
A method for removing periodic noise from a reconstructed light field image includes the steps of: acquiring a light field image of a sample; acquiring an optical center position map without the sample; calibrating the imaging centers of the microlenses and performing reconstruction on the light field image; transforming a reconstructed light field image to the frequency domain and generating an image frequency spectrum; preprocessing the image frequency spectrum; generating a low-pass filter; multiplying the low-pass filter with the preprocessed image frequency spectrum, and then setting the frequency spectrum value of the low-frequency component to zero; performing binarization on the reconstructed light field image frequency spectrum to obtain an image mask; removing the high-frequency periodic noise component from the original frequency spectrum of the reconstructed light field image; and transforming the filtered reconstructed light field image frequency spectrum back to the spatial domain to obtain the reconstructed light field image.