Multi-Resolution Image Denoising With Edge-Aware Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing denoising methods for image sensors result in blurry images with insufficient edge preservation, leading to noise at strong edges and aliasing, necessitating subsequent sharpening algorithms.
Innovation Solution
A denoising method that involves filtering residual images at different resolutions, enhancing weak edges, attenuating strong edges, and reconstructing images to generate a target image, thereby increasing local contrast and edge details without additional sharpening.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing denoising algorithms are applied to image sensor output, then noise is reduced and signal-to-noise ratio is improved, but image definition deteriorates and images become blurry
Solution Approach 1:
The patent segments the image into multiple layers based on edge strength (strong edges, weak edges, and non-edge regions) and applies different processing strategies to each layer. Strong edge regions are preserved with minimal filtering to maintain definition, while non-edge regions undergo aggressive denoising. This layered segmentation resolves the contradiction by allowing simultaneous noise reduction and definition preservation through region-specific processing.
2Manufacturing precision
If edge-preserving denoising algorithms are used, then image definition is maintained, but denoising strength at strong edges is insufficient and noise remains at edges causing aliasing
Solution Approach 1:
The patent applies local quality by assigning different denoising strengths to different spatial regions based on edge characteristics. Strong edge regions receive minimal filtering (preserving definition but allowing some noise), weak edge regions receive moderate filtering, and non-edge regions receive maximum filtering. This local differentiation resolves the contradiction by optimizing denoising strength for each region's specific requirements rather than applying uniform processing.
3Manufacturing precision
If denoising is performed first and then sharpening is applied to improve definition, then image clarity is improved, but the implementation process becomes complex with multiple steps
Solution Approach 1:
The patent merges the denoising and sharpening operations into a single integrated process. By segmenting the image into edge strength layers and applying appropriate filtering to each layer during the denoising phase, the algorithm simultaneously achieves noise reduction and edge enhancement. This consolidation resolves the contradiction by eliminating the need for separate sharpening steps while maintaining image clarity through the layered processing approach.
Data Source
Figure 1~2
Figure 3~4
AI summary
The present disclosure relates to a denoising method, a denoising apparatus, an electronic device and a computer-readable storage medium, wherein the method includes: performing filtering on a residual image of an image generate a filtered image, wherein the residual image is a difference value of the image at different resolutions, the image is corresponding to at least two residual images, and the at least two residual images have a preset layer sequence relationship therebetween; performing enhancement on a weak edge region in the filtered image, and performing attenuation on a strong edge region in the filtered image, so as to generate a sharpened image; and obtaining a target image according to a reconstructed image and the sharpened image, wherein the reconstructed image is superposition of all residual images whose layer sequence is before the residual image. Sharpening is performed during the superposition of reconstruction, thus increasing local contrast, and enhancing edges and details, so that sharpening does not need to be performed after the image reconstruction. In addition, performing attenuation on the strong edge region may reduce noise in the strong edge region, and reduce glitches of the image.