RAW Low-Light Image Denoising Against Dead Pixel Spread
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image denoising methods fail to effectively suppress the influence of dead pixels in low-light images, leading to reduced image quality and loss of edge details due to the spread of highlighted dead pixels through convolutional neural networks.
Innovation Solution
A method involving image enhancement transformations followed by a pre-trained denoising network model, trained on low-light images with simulated dead pixels, to automatically suppress dead pixels and enhance dark details, using normalization and gamma transformations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional image processing methods are used to correct dead pixels, then dead pixel correction is attempted, but edge details of the image are damaged due to smooth effect similar to median filtering
Solution Approach 1:
The patent applies image enhancement transformation (gamma correction and normalization) to the low-light image before inputting it to the denoising network model. This preliminary action transforms the image data to a state where dead pixels are less prominent and the network model can process them more effectively, preventing the spread of dead pixel influence while preserving edge details.
Solution Approach 2:
The patent introduces a pre-trained denoising network model as an intermediary between the raw low-light image and the final denoised output. This network model, trained on synthetic data with simulated dead pixels, acts as a mediator that automatically suppresses dead pixels through learned features rather than direct pixel manipulation, thereby preserving edge details while correcting dead pixels.
2Loss of energy
If deep learning denoising is applied in RAW domain, then denoising is performed, but dead pixels spread from a single pixel to several pixels or even dozens of pixels due to receptive-field mechanism
Solution Approach 1:
The patent performs image enhancement transformation (gamma correction with gamma=0.5 and normalization) on the low-light image before inputting it to the denoising network. This preliminary transformation reduces the prominence of dead pixels in the input data, preventing their harmful influence from being amplified by the network's receptive field mechanism during the denoising process.
Solution Approach 2:
The patent changes the parameter representation of the image by applying gamma correction and normalization transformations. This parameter change modifies the distribution of pixel values, making dead pixels less dominant in the transformed space, thereby preventing their influence from spreading when processed by the convolutional neural network.
3Loss of energy
If conventional denoising methods are used on low-light images, then noise is reduced, but image quality is reduced due to inability to suppress highlighted dead pixels
Solution Approach 1:
The patent applies image enhancement transformation (gamma correction and normalization) as a preliminary step before denoising. This transformation enhances dark details and reduces the prominence of dead pixels, allowing the subsequent denoising network to effectively reduce noise while maintaining image quality by automatically suppressing the transformed dead pixel artifacts.
Solution Approach 2:
The patent introduces a pre-trained denoising network model as an intermediary that has been specifically trained on synthetic low-light images with simulated dead pixels. This network learns to distinguish between noise and dead pixel artifacts, enabling it to reduce noise while automatically suppressing dead pixels and preserving image quality.
Data Source
AI summary
Examples of the present disclosure provide a method and an apparatus for denoising a low-light image, wherein the method includes: acquiring a low-light image in RAW domain; performing a preset image enhancement transformation on the low-light image in the RAW domain, and inputting the transformed image into a pre-trained denoising network model to obtain an output image; wherein, the denoising network model is trained based on sample images, and the sample images include a low-light image with simulated dead pixels and a noiseless image; performing an inverse transformation of the preset image enhancement transformation on the output image to obtain a denoised image. It can significantly reduce the influence of dead pixels of images on the process of denoising the low-light image and improve the quality of low-light image denoising.