Image Super-Resolution via Median Random Smoothing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution methods using deep neural networks struggle to generalize to real-world images and are not robust against various disturbances such as noise and adversarial attacks, often requiring computationally expensive training with adversarial examples that provide limited protection against specific types of perturbations.
Innovation Solution
A method for image super-resolution using a neural network that incorporates a median random smoothing inference phase, generating multiple noisy images with Gaussian noise and averaging their predictions to enhance robustness, combined with a training phase that adjusts the network using median random smoothing and diverse noise levels to improve generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If adversarial training is used to improve robustness against specific types of perturbations, then reliability improves, but device complexity and computational cost increase
Solution Approach 1:
The patent changes the parameter of noise standard deviation to multiple predefined values during training, allowing the model to learn robustness across different noise levels without requiring complex adversarial training procedures for each attack type
Solution Approach 2:
The trained model achieves universal robustness against multiple types of disturbances including Gaussian noise, salt-and-pepper noise, and adversarial attacks, functioning as a multi-purpose robust model without requiring separate training for each attack type
2Manufacturing precision
If multiple neural network models are created and trained to recover noise and details lost during bicubic downsampling, then manufacturing precision improves, but productivity decreases due to computational expense
Solution Approach 1:
The patent performs preliminary training of the neural network on clean image pairs before deployment, so that during actual use, the pre-trained model can directly process noisy images without requiring complex runtime processing or multiple models
Solution Approach 2:
The model is trained with images corrupted by Gaussian noise with different standard deviations, enabling it to handle various noise levels during inference without requiring retraining or multiple specialized models
3Manufacturing precision
If a neural network is trained on bicubically downsampled images, then manufacturing precision improves for synthetic data, but adaptability decreases for real-world images with different distortion types
Solution Approach 1:
The patent introduces variation in noise standard deviation parameters during training, exposing the model to diverse corruption patterns that mimic real-world imaging conditions, thereby improving generalization to unseen image types and distortions
Solution Approach 2:
The method skips the intermediate step of creating multiple specialized models for different image types by training a single model on diverse noisy data, allowing direct deployment to various real-world scenarios
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The invention relates to a method for image super-resolution. The method comprises a specific inference phase (300) carried out from a neural network (13) previously trained to produce a high-resolution image from a low-resolution image. The inference phase (300) comprises: - a generation (310) of a plurality of noisy images from a low-resolution image, each noisy image being obtained by a random application to the low-resolution image of a Gaussian noise having a predefined standard deviation, - a prediction (320), by the previously trained neural network, of an image for each of the noisy images thus obtained, - a generation (330) of a high-resolution image corresponding to a median of the predicted images thus obtained.