Image Processing Model Training Using Real Degradation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current super-resolution algorithms are ineffective in processing real images with noise and blurriness, as they rely on bicubic operators to construct training data that does not reflect real image degradation, and fail to adequately address blurriness and noise issues in low-resolution images.
Innovation Solution
A novel training method for an image processing model that uses unlabeled real images to train a neural network, constructing training image pairs that retain noise and blurriness features, allowing the network to generate sharper and cleaner high-resolution images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If bicubic operators are used to construct training data, then the training process is simple and stable, but the low-resolution images do not reflect real degradation scenarios with noise and blurriness
Solution Approach 1:
The patent changes the parameter of image degradation by using real degradation models instead of fixed bicubic operators. The system dynamically adjusts degradation parameters based on actual image characteristics, allowing the training data to reflect real-world noise and blurriness patterns while maintaining computational feasibility through automated degradation simulation.
Solution Approach 2:
The patent creates synthetic low-resolution images by copying and degrading high-resolution images through real degradation models. This copying process generates training data that mirrors actual image degradation scenarios, enabling the neural network to learn from realistic degradation patterns rather than idealized bicubic downsampling.
2Stability of the object's composition
If high-resolution images are downscaled using bicubic operators, then the training process is stable, but the generated low-resolution images lack real noise and blurriness features
Solution Approach 1:
The patent converts the harmful effect of image degradation into a beneficial training feature. Instead of avoiding noise and blurriness in training data, the system intentionally introduces real degradation models to create training images that contain these features. This allows the neural network to learn effective denoising and deblurring capabilities while maintaining training stability through automated degradation simulation.
3Productivity
If current super-resolution algorithms are used, then the processing is fast and stable on clean images, but the texture details are insufficient for reconstructing realistic textures
Solution Approach 1:
The patent changes the training parameters by using real degradation models that capture complex noise and blurriness characteristics. This enables the neural network to learn more accurate texture reconstruction patterns while maintaining efficient processing speeds through optimized training data generation that reflects actual image degradation physics.
Data Source
AI summary
A method, an apparatus, and a device for image processing and a training method thereof are provided. The training method includes obtaining a sample image set, the sample image set comprising a first number of sample images; constructing an image feature set based on the sample image set, the image feature set comprising an image feature extracted from each of the sample images in the sample image set; obtaining a training image set, the training image set comprising a second number of training images; constructing multiple training image pairs based on the training image set and the image feature set; and training the image processing model based on the multiple training image pairs.