Invertible Neural Networks for Unknown-Downsampling Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning-based super-resolution methods rely on paired high-resolution and low-resolution images, which are often artificially synthesized, leading to poor performance when processing low-resolution images generated through unknown downsampling methods.
Innovation Solution
An invertible neural network is trained to generate intermediate images and high-frequency information based on input images, allowing for the reconstruction of high-resolution outputs using an inverse network, simulating unknown downsampling processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learning methods use artificially synthesized image pairs for training, then the training process can be completed, but the model performance deteriorates when processing low-resolution images from unknown downsampling methods
Solution Approach 1:
The patent inverts the traditional training approach by training the model to generate low-resolution images from high-resolution images (downsampling task) rather than generating high-resolution images from low-resolution images. This inversion allows the model to learn the downsampling process itself, making it adaptable to unknown downsampling methods when applied to super-resolution tasks.
Solution Approach 2:
The patent changes the fundamental parameter of the learning task from super-resolution (low-to-high resolution) to downsampling (high-to-low resolution). By changing this key parameter, the model learns a more generalizable representation that adapts to various unknown downsampling methods rather than being optimized for a specific synthetic downsampling process.
2Ease of manufacture
If paired high-resolution and low-resolution images are used for training, then the machine learning model can be trained, but the training data becomes artificially synthesized rather than representing real-world scenarios
Solution Approach 1:
Instead of synthesizing low-resolution images from high-resolution images for training (traditional approach), the patent inverts the process by synthesizing high-resolution images from low-resolution images during the training phase. This allows the model to learn from a more realistic distribution of image pairs while maintaining ease of data preparation.
3Productivity
If the model is trained on low-resolution images from specific downsampling methods, then training convergence is achieved, but the model cannot effectively process low-resolution images from other downsampling methods
Solution Approach 1:
The patent makes the model universal by training it to learn the downsampling process itself rather than inverting a specific downsampling method. This multi-functional approach allows the model to handle various unknown downsampling methods (bicubic, bilinear, nearest neighbor, etc.) with a single trained model, achieving both training convergence and broad generalization.
Data Source
AI summary
According to implementations of the subject matter described herein, a solution is proposed for super-resolution image reconstructing. According to the solution, an input image with first resolution is obtained. An invertible neural network is trained using the input image, wherein the invertible neural network is configured to generate an intermediate image with second resolution and first high-frequency information based on the input image, the second resolution being lower than the first resolution. Subsequently, an output image with third resolution is generated based on the input image and second high-frequency information by using an inverse network of the trained invertible neural network, the second high-frequency information conforming to a predetermined distribution, and the third resolution being higher than the first resolution. The solution can effectively process a low-resolution image obtained by an unknown downsampling method, thereby obtaining a high-quality and high-resolution image.