Wavelet-Domain Normalizing Flow for Stable Image Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image super-resolution reconstruction methods, particularly those based on deep learning and normalizing flow models, suffer from instability and inefficiency in generating high-quality super-resolution images, as they often produce multiple unsatisfactory outcomes and lack stability during training.
Innovation Solution
A method and system for wavelet domain-based normalizing flow super-resolution image reconstruction, utilizing a stable normalizing flow model constructed with wavelet transforms, QR layers, T-distributions, and refinement layers to enhance image reconstruction quality and stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning models with fixed parameters are used for super-resolution, then the model structure is simple and easy to train, but the model cannot capture the multiple possible high-resolution images corresponding to a single low-resolution image, leading to limited generative capability
Solution Approach 1:
The patent changes the parameter representation from fixed deterministic values to probabilistic distributions. The normalizing flow model learns to map from a simple prior distribution to a complex posterior distribution of possible high-resolution images, allowing multiple valid solutions while maintaining a structured approach to the generative process
Solution Approach 2:
The patent introduces a normalizing flow model as an intermediary between the low-resolution input and high-resolution output. This intermediary learns the complex mapping relationship through a series of invertible transformations, enabling the model to generate multiple plausible high-resolution images while maintaining computational tractability
2Adaptability or versatility
If normalizing flow model is used for super-resolution, then the model can generate multiple super-resolution images with different details, but the training process becomes unstable
Solution Approach 1:
The patent changes the distribution assumption from Gaussian to T-distribution with heavy tails. This parameter change in the probability distribution makes the training objective more robust to outliers and numerical instabilities, thereby improving training stability while preserving the generative capability
Solution Approach 2:
The patent implements a feedback mechanism where the model generates multiple candidate high-resolution images and uses a discriminator to provide feedback on their quality. This adversarial feedback guides the normalizing flow to produce more realistic and diverse images while stabilizing the training process through the game-theoretic equilibrium
3Adaptability or versatility
If deterministic mapping is used in super-resolution models, then the model is simple to implement, but it cannot represent the ambiguity inherent in the inverse problem where one low-resolution image corresponds to multiple high-resolution images
Solution Approach 1:
The patent transforms the deterministic parameter mapping into a probabilistic distribution mapping. Instead of learning a single fixed transformation, the normalizing flow learns to map to a distribution over possible transformations, naturally representing the ambiguity in the inverse problem while maintaining a structured and computationally efficient framework
Data Source
AI summary
The present disclosure discloses a method and a system for wavelet domain-based normalizing flow super-resolution image reconstruction. The method includes constructing a training set and a normalizing flow model, wherein the normalizing flow model includes a plurality of levels, each of the plurality of levels including a squeeze layer, two types of conditional mapping layers, a split layer, an activation standard layer, and a quick response (QR) layer; determining a stable normalizing flow model through a wavelet transform, a reconstructed QR layer, and a T-distribution based on the normalizing flow model; determining a wavelet domain-based normalizing flow super-resolution model by adding a refinement layer based on the stable normalizing flow model; training the wavelet domain-based normalizing flow super-resolution model based on the training set; and reconstructing a super-resolution image based on a trained normalizing flow super-resolution model.