Image Enhancement Using Non-Normal Noise Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based image enhancement methods struggle with accurately handling real image noise that does not follow a normal distribution, due to the difficulty in obtaining sufficient training pairs of noisy and clean images and pixel position misalignments, leading to less accurate training and inference operations.
Innovation Solution
A processor-implemented method generates a non-normal noise map using a noise model based on non-normal noise distribution information, including pixel position and image capturing parameters, and enhances images using an image enhancement model trained on these noise maps, addressing the non-normal noise distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If deep learning-based neural network is used for image enhancement with presumption of normal distribution noise, then the training and inference operations can be performed with standard methods, but the accuracy deteriorates when dealing with real image noise that does not follow normal distribution
Solution Approach 1:
The patent changes the fundamental parameter assumption from normal distribution to non-normal distribution (e.g., Poisson, Gamma, or custom distributions) to match real image noise characteristics. This allows the noise model to accurately represent actual noise while still enabling systematic training through distribution-based generation of training pairs.
Solution Approach 2:
The patent creates synthetic training pairs by copying and transforming clean images through the non-normal noise model to generate corresponding noisy images. This copying approach generates sufficient training data without requiring extensive real noisy-clean image pairs, solving the data scarcity problem while maintaining accuracy.
2Measurement precision
If sufficient training pairs of noisy and clean images are obtained for real noise, then the training accuracy can be improved, but the difficulty increases due to pixel position misalignments and data collection challenges
Solution Approach 1:
The patent introduces a noise model as an intermediary between clean images and noisy images. This mediator systematically generates noisy versions of clean images with known ground truth, eliminating the need to manually collect and align real noisy-clean pairs. The intermediary approach simplifies the training data preparation process while ensuring pixel-perfect alignment.
Solution Approach 2:
The patent performs preliminary noise modeling and training pair generation before the actual training process. By pre-processing the data through the non-normal noise model to create aligned training pairs, the system eliminates the complexity of real-time noise handling during training and ensures optimal training conditions from the start.
3Ease of operation
If standard noise models assuming normal distribution are used, then the computational process can be simplified, but the image enhancement accuracy deteriorates for real images with non-normal noise characteristics
Solution Approach 1:
The patent implements dynamic noise modeling where the noise distribution parameters are adapted to match the specific characteristics of each input image or scene. This dynamic approach allows the system to maintain computational efficiency through parameterized models while achieving high accuracy by adjusting to non-normal noise characteristics specific to different imaging conditions.
Solution Approach 2:
The patent segments the image processing into distinct stages: noise estimation, distribution identification, and enhancement processing. This segmentation allows each stage to be optimized independently, maintaining computational simplicity in the noise modeling stage while achieving high accuracy in the enhancement stage through non-normal distribution handling.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method and apparatus with noise consideration are provided. The method includes generating, using a noise model, a non-normal noise map corresponding to a noise of an input image, and generating an enhanced image of the input image by implementing an image enhancement model based on the input image and the non-normal noise map, where the noise of the input image follows a non-normal distribution.