Non-local Adaptive Loop Filter for Video Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression technologies introduce compression noise due to coarse quantization and motion compensation, leading to artifacts like blocking, ringing, and blurring in reconstructed pictures, which current in-loop filters struggle to fully mitigate.
Innovation Solution
A non-local adaptive loop filter (NL-ALF) method that divides reconstructed pictures into patches, forms patch groups with similar reference patches, and applies denoising techniques using a compression noise model to modify pixel values, with adaptive on/off control flags for each block to determine whether to adopt filtered or original pixel values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If in-loop filters are applied to reduce compression noise, then picture quality is improved, but device complexity increases
Solution Approach 1:
The picture is divided into multiple patches, and each patch is processed independently to form patch groups. This segmentation allows the filter to focus on local regions with similar characteristics, improving denoising effectiveness while managing computational complexity through localized processing.
Solution Approach 2:
The filter adapts its behavior to local picture characteristics by forming patch groups based on similarity metrics. Each patch group is processed with parameters optimized for its specific content type (e.g., smooth regions vs. textured regions), achieving better quality results without uniformly increasing complexity across the entire picture.
2Loss of information
If non-local adaptive loop filter is applied to reduce compression noise, then bit rate is reduced, but loss of information increases
Solution Approach 1:
The filter uses a feedback mechanism where patch groups are formed based on similarity to the current patch, and the filtering operation is adaptively applied based on the characteristics of these groups. This feedback-driven approach ensures that denoising is applied selectively, reducing compression artifacts while preserving important picture information and maintaining quality.
Solution Approach 2:
The filter dynamically adjusts its parameters based on local picture characteristics. By changing filtering parameters according to the specific content of each patch group (e.g., noise variance, similarity metrics), the system achieves effective noise reduction without over-filtering and losing important picture details.
3Manufacturing precision
If patch groups are formed with multiple reference patches, then denoising effectiveness is improved, but processing time increases
Solution Approach 1:
The picture is segmented into patches and organized into patch groups, allowing parallel processing of multiple regions. This segmentation enables the system to process multiple patches simultaneously, reducing overall processing time while maintaining effective denoising through the use of multiple reference patches within each group.
Solution Approach 2:
Patch groups are pre-formed by identifying similar reference patches before the actual denoising operation. This preliminary organization of patches into groups based on similarity metrics allows the filtering stage to proceed more efficiently, as the candidate reference patches have already been identified and organized, reducing the computational burden during the actual denoising process.
Data Source
AI summary
Aspects of the disclosure provide a method for denoising a reconstructed picture. The method can include receiving reconstructed video data corresponding to a picture, dividing the picture into current patches, forming patch groups each including a current patch and a number of reference patches that are similar to the current patch, denoising the patch groups to modify pixel values of the patch groups to create a filtered picture, and generating a reference picture based on the filtered picture for encoding or decoding a picture. The operation of denoising the patch groups includes deriving a variance of compression noise in the respective patch group based on a compression noise model. A selection of model parameters is determined based on coding unit level information.