Bayer Domain Noise Filtering for Digital Cameras
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image noise filtering methods in digital cameras are either expensive to implement, lead to loss of high-frequency image parts, or result in color-bleeding artifacts, especially when dealing with low-frequency noise in high-resolution images and chroma channels.
Innovation Solution
Applying a noise filter with small support directly in the Bayer domain to suppress strong chrominance noise without damaging luminance detail, equivalent to applying strong filtering to chrominance channels and weak filtering to luminance channels in the color space domain, without requiring conversion, and using a method that divides images into sub-images for filtering to preserve high-frequency components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a spatial filter with large spatial support is applied to remove low-frequency noise, then noise reduction effectiveness is improved, but implementation cost increases
Solution Approach 1:
The patent segments the image into multiple blocks and processes each block independently with a small-support filter. This segmentation allows the use of computationally inexpensive small filters while achieving noise reduction across the entire image, resolving the contradiction between noise reduction effectiveness and implementation cost.
Solution Approach 2:
The patent transforms the filtering operation from spatial domain to frequency domain using DCT (Discrete Cosine Transform). By operating in the frequency domain, a small spatial support filter achieves low-frequency noise removal equivalent to large spatial support filtering, reducing computational complexity while maintaining effectiveness.
2Reliability
If downsampling is applied to reduce resolution before filtering, then low-frequency noise reduction is improved, but high-frequency image details are lost
Solution Approach 1:
The patent applies DCT transformation to operate in the frequency domain, allowing selective filtering of low-frequency noise components while preserving high-frequency details. This dimensional transformation enables differential treatment of frequency components without the information loss inherent in spatial downsampling.
Solution Approach 2:
The patent applies different filtering strengths to different frequency components through DCT-based filtering. Low-frequency components receive strong filtering to remove noise, while high-frequency components are preserved to maintain image details, achieving local quality differentiation in the frequency domain.
3Reliability
If a strong noise filter is applied in the Bayer domain, then chrominance noise reduction is improved, but luminance detail is lost
Solution Approach 1:
The patent applies different filtering strengths to different color channels in the Bayer domain. The blue channel (which contains primarily chrominance information) receives strong filtering to suppress noise, while the green channel (which contains luminance information) receives weak or no filtering to preserve details. This local quality differentiation resolves the contradiction between noise reduction and detail preservation.
Solution Approach 2:
The patent applies filtering selectively to specific channels rather than uniformly to all channels. By applying strong filtering only to the blue channel where chrominance noise predominates, and minimal filtering to the green channel where luminance details exist, the patent achieves partial action that targets noise without damaging important image information.
4Reliability
If conversion from Bayer domain to color space domain is performed for separate filtering, then selective noise filtering is improved, but computational complexity increases
Solution Approach 1:
The patent extracts and processes only the blue channel (which contains the problematic chrominance noise) separately with strong filtering, while leaving the green and red channels minimally processed. This selective extraction approach achieves selective noise filtering without the computational overhead of converting to color space domain and processing all channels separately.
Solution Approach 2:
The patent uses a simple, computationally inexpensive approach by applying different filtering strengths directly to Bayer domain channels without color space conversion. This disposable-like simplicity in the filtering approach achieves the desired selective noise reduction without the expensive computational operations of full color space transformation.
Data Source
AI summary
Methods and systems for image noise filtering are provided. One method of image noise filtering includes generating a plurality of sub-images of a digital image, applying a noise filter with small support to each sub-image, and generating a filtered digital image by merging the filtered sub-images. Another method of image noise filtering includes receiving a digital image including a plurality of color channels in the Bayer domain, applying a strong noise filter to each color channel to generate filtered color channels, computing a luminance image from the digital image, applying a weak noise filter to the luminance image to generate a filtered luminance image, computing a luminance recovery factor map using the luminance image and the filtered luminance image, and computing output color channels of the digital image using the filtered color channels and the luminance recovery factor map.