Dual-Model Noise Estimation for Image Sharpness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for estimating noise in images, particularly in low-light conditions, are inadequate, as they either blur images or fail to accurately remove noise, leading to suboptimal image quality.
Innovation Solution
A dual-model approach is employed, where one model is trained to detect stochastic noise (Poisson and Gaussian) and another to detect defective pixel noise, with both models processing images through stages of resolution reduction and increase, using skip connections for effective noise estimation and de-noising.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a filter is applied to blur the image to reduce high-frequency noise, then noise reduction is improved, but image sharpness deteriorates
Solution Approach 1:
The patent segments the noise estimation process into two separate trained models: one for detecting stochastic noise and another for detecting extreme pixel values (defective pixel noise). Each model processes the image independently through resolution reduction stages with skip connections, allowing specialized noise removal without affecting image sharpness. The segmented approach enables selective noise removal while preserving structural details.
Solution Approach 2:
The patent introduces an intermediary noise estimation process that occurs before the actual de-noising operation. By first estimating the noise characteristics using the dual models and then using this estimation to guide the de-noising process, the system can remove noise adaptively without applying excessive blurring. The noise estimate serves as a mediator between the noisy input and the cleaned output.
2Illumination intensity
If sensor gain is increased to amplify brightness in low light conditions, then image brightness is improved, but stochastic noise increases
Solution Approach 1:
The patent implements feedback through the skip connections in the neural network architecture, where features from earlier resolution stages are fed back to later stages. This allows the model to learn the relationship between brightness and noise characteristics, enabling it to distinguish between valid signal and stochastic noise even in high-gain, low-light images. The feedback mechanism helps the model adapt to varying noise levels without requiring manual gain adjustment.
Solution Approach 2:
The patent changes the processing parameters by reducing image resolution through multiple stages while preserving essential features through skip connections. This parameter transformation allows the model to detect noise patterns that are not visible at full resolution, enabling more accurate noise estimation and removal in low-light conditions without amplifying stochastic noise.
3Device complexity
If a single model is used for noise estimation, then device complexity is reduced, but noise estimation accuracy deteriorates
Solution Approach 1:
The patent divides the noise estimation task into two separate specialized models: one trained specifically for stochastic noise detection and another for defective pixel noise detection. Each model has its own set of parameters and architecture optimized for its specific noise type. This segmentation improves accuracy by allowing each model to specialize in detecting its target noise pattern without interference from other noise types.
Solution Approach 2:
The patent merges the outputs of the two separate noise estimation models to create a comprehensive noise estimate. By combining the stochastic noise estimate and the defective pixel noise estimate, the system achieves superior overall noise removal performance. The merging process integrates the strengths of both specialized models while maintaining manageable complexity through modular architecture.
Data Source
Figure 1~3
Figure 4~6
Figure 5
AI summary
An image processing device comprising a processor configured to estimate noise in an image, the image being represented by a set of pixels and each pixel having a value associated with it on each of one or more channels, by the steps of: processing data derived from the image by means of a first trained model for detecting stochastic noise so as to form a first noise estimate; processing data derived from the image by means of a second trained model for detecting extreme pixel values to form a second noise estimate; and combining the first and second noise estimates to form an aggregated noise estimate.