CNN Image Denoising for Motion-Induced Ghosting Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image and video denoising methods using multi-frame averaging suffer from ghosting artifacts due to device shake or object motion during image capture.
Innovation Solution
A denoising method utilizing a convolutional neural network trained with a preset training set that includes motion-simulated noiseless images and noisy counterparts, simulating various capture scenarios to learn effective denoising without ghosting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-frame averaging method is used for denoising, then noise reduction effect is improved, but ghosting artifacts appear due to device shake or object motion
Solution Approach 1:
The patent transforms the denoising problem from simple averaging to a parameter optimization problem by using a convolutional neural network. The network learns optimal denoising parameters through training on motion-simulated data, enabling it to adaptively adjust denoising strength and preserve moving objects while removing noise, thus resolving the contradiction between noise reduction and ghosting prevention
Solution Approach 2:
The patent applies preliminary action by pre-training the convolutional neural network with motion-simulated training sets that include device shake and object motion scenarios. This preliminary training enables the model to learn how to handle motion artifacts before actual denoising, allowing it to preserve moving objects while removing noise without producing ghosting artifacts
Data Source
AI summary
The embodiments of the present application provide a denoising method, apparatus, electronic device and medium, which relates to the technical field of image processing. The method comprises: inputting an image to be processed into an image denoising model, which is a model obtained by training a convolutional neural network model based on a preset training set, wherein the preset training set includes multiple groups of annotation data and sample data corresponding to each group of annotation data, and each group of annotation data includes multiple noiseless images obtained by performing motion simulation processing on one reference noiseless image, and the sample data corresponding to the group of annotation data includes images obtained by superimposing noise to the multiple noiseless images respectively; acquiring denoised image data output by the image denoising model; converting the image data into an image, to obtain a denoised image corresponding to the image to be processed. This can effectively solve the problem of ghost existed in the denoised image.