Image Denoising via Unpaired Noise Generation Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current super-resolution algorithms trained on paired data sets often fail to produce effective results when applied to low-resolution images with distributions different from the training data, leading to artifacts and poor image quality.
Innovation Solution
An image denoising method that uses unpaired training sets to train a noise generation model and an image denoising network, where the noise generation network and discrimination network are alternately trained to generate simulated noise and adjust parameters, improving the image super-resolution effect by converting low-resolution images into high-resolution images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If paired training sets are used for super-resolution algorithm training, then the training process is simple and efficient, but the algorithm produces artifacts and poor image quality when applied to images with distributions different from training data
Solution Approach 1:
The patent introduces a noise generation network as an intermediary component that bridges the gap between clean low-resolution images and noisy high-resolution images. This network generates realistic noise patterns that are then added to training images, allowing the model to learn robust denoising capabilities without requiring actual paired noisy-clean image datasets. The intermediary noise generation mechanism enables the system to handle distribution differences effectively.
Solution Approach 2:
The patent performs preliminary noise generation and addition to training images before the main super-resolution training process. By pre-processing the training data with synthesized noise, the model is prepared in advance to handle noisy inputs with different distributions, improving its generalization capability and reducing artifacts in final output images.
2Reliability
If unpaired training sets are used to train the noise generation model and image denoising network, then image quality and definition are improved, but the training process becomes more complex
Solution Approach 1:
The patent segments the training process into distinct phases: first training a noise generation network separately, then using its output to train the image denoising network. This segmentation allows each network to be optimized independently with appropriate loss functions and training strategies, managing complexity through modular approach while achieving superior image quality.
Solution Approach 2:
Instead of directly training a denoising network on unpaired data, the patent inverts the approach by first training a noise generation network to create synthetic noisy images, then using these generated images for denoising training. This inverted approach simplifies the overall training process by breaking down the complex unpaired learning problem into two more manageable supervised learning tasks.
3Reliability
If high-resolution image sensors are used to obtain high-resolution images, then image detail and quality are improved, but manufacturing cost and process limitations make this method difficult to implement
Solution Approach 1:
The patent replaces the mechanical approach of using high-resolution image sensors with a computational approach using deep learning-based super-resolution algorithms. Instead of relying on expensive hardware upgrades, the system uses software-based image processing to reconstruct high-resolution images from low-resolution inputs, significantly reducing manufacturing costs and implementation complexity while maintaining image quality.
Data Source
AI summary
An image denoising method includes: acquiring a first data set and a second data set, where the first data set includes a plurality of first images without noise, the second data set includes a plurality of second images with real noise, contents of each first image and each second image are different; training, by using the first data set and the second data set, a first network to obtain a noise generation model; inputting the first image into the noise generation model, and outputting a third image with simulated noise; where a plurality of third images forms a third data set; training, by using the first data set and the third data set, an image denoising network to obtain an image denoising model. The image denoising model is configured to convert an original image with noise into an output image without noise.