RAW Image Noise Reduction Using Color Filter-Aware Non-Local Means
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional noise reduction methods for RAW image data, such as those using low-pass filters or adaptive noise reduction techniques, struggle to effectively reduce noise while preserving image quality, especially when pixels have signal values of different levels due to color filters, leading to image degradation.
Innovation Solution
An image processing apparatus and method that determines a target pixel group and reference pixel group based on color filter array information, using the Non-local Means algorithm to calculate weighted averages for noise reduction, ensuring accurate noise reduction even in images with varying pixel signal levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If a low-pass filter is used to reduce noise, then noise is reduced, but edge sharpness is blurred
Solution Approach 1:
The patent applies different processing strategies to different regions of the image based on local characteristics. By calculating gradient magnitudes and variances for each pixel, the algorithm identifies edge regions and applies noise reduction only to non-edge areas, preserving edge sharpness while reducing noise in homogeneous regions.
Solution Approach 2:
The patent uses dynamic thresholding where the noise reduction strength varies pixel-by-pixel based on local image characteristics. The gradient magnitude and variance calculations create spatially-varying weights that adapt to local edge density and texture, allowing the filter to be aggressive in smooth areas and conservative near edges.
2Object-affected harmful factors
If adaptive noise reduction is applied to RAW image data with different color filter signal levels, then noise reduction is achieved, but image quality degrades due to incorrect reference pixel selection
Solution Approach 1:
The patent transforms the RAW image data by converting it to a luminance-based representation where pixel values reflect brightness rather than color-filter-specific signal levels. This parameter transformation allows standard adaptive noise reduction algorithms to work correctly by comparing luminance similarity rather than raw signal levels, enabling proper reference pixel selection across the entire image.
3Illumination intensity
If ISO sensitivity is increased to capture more light, then image brightness is improved, but noise is significantly increased
Solution Approach 1:
The patent creates multiple copies of pixel values from similar regions in the image to build a distribution for noise estimation. By finding reference pixels with similar luminance values and copying their statistical properties, the algorithm can distinguish between noise and actual signal, enabling effective noise reduction even in low-light conditions where ISO must be increased.
Data Source
AI summary
Noise in RAW image data is reduced by determining parameters including pixels used as a target area and reference pixels in the RAW image data, based on color filter information for the RAW image data. The RAW image data is corrected based on the parameters thus determined.