Neural Network Gamma Correction for RAW Image Estimation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods using multilayer neural networks struggle to maintain estimation accuracy when dealing with RAW images, as the error calculation is influenced by gamma correction, leading to issues like undershoot and ringing, especially in high-resolution upsampling and deblurring, and high-contrast dehazing.
Innovation Solution
The method involves performing gamma correction on both the correct and output images during the learning step to calculate errors, updating the network parameters based on these corrected values, thereby maintaining a linear relationship between light amount and signal value, and suppressing gradation collapse in high luminance parts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If gamma correction is not applied during error calculation in neural network learning, then the learning process is simpler and faster, but the estimation accuracy deteriorates in dark parts of the developed image
Solution Approach 1:
The patent applies gamma correction to the correct image and output image before calculating the error, preparing the data in advance to match the actual display conditions. This preliminary transformation ensures that the error calculation reflects the true visual quality, preventing the network from learning incorrect patterns that would otherwise occur when using uncorrected linear values.
Solution Approach 2:
The patent transforms the error calculation from using linear signal values to using gamma-corrected values. By changing the parameter space in which the error is computed, the learning process becomes aligned with human visual perception and actual image display characteristics, thereby improving estimation accuracy without significantly increasing computational burden.
2Measurement precision
If gamma correction is applied during error calculation, then the estimation accuracy is maintained across different signal magnitudes, but the calculation load increases
Solution Approach 1:
The gamma correction is applied in advance to the correct and output images before error calculation. This preliminary transformation allows the network to learn from errors that already account for gamma effects, eliminating the need for repeated gamma corrections during backpropagation and reducing the overall computational load despite the initial transformation cost.
3Ease of manufacture
If the network learns using linear RAW image values, then the learning process is straightforward, but the error grows in dark parts of the estimated image after gamma correction
Solution Approach 1:
The patent changes the parameter space for error calculation from linear RAW values to gamma-corrected values. This transformation ensures that errors in dark regions are weighted appropriately according to human visual sensitivity, preventing the network from prioritizing bright regions and improving overall estimation accuracy across all luminance levels.
Data Source
AI summary
An image processing method includes the steps of acquiring a training image and a correct image, inputting the training image into a multilayer neural network to generate an output image, performing a gamma correction for each of the correct image and the output image and calculating an error between the correct image after the gamma correction and the output image after the gamma correction, and updating a network parameter of the neural network using the error.