Deep Perceptual Image Enhancement Network for Exposure Uniformity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image enhancement technologies face challenges in achieving global uniformity for varying exposure inputs, preserving high-frequency content, and reducing artifacts such as halation and noise, especially in deep learning methods trained for standard or low exposure images.
Innovation Solution
The proposed solution involves a deep perceptual image enhancement network (DPIENet) with a neural input enhancer (NIE) that uses logarithmic exposure transformation and a dynamic channel attention mechanism, along with a multi-scale human vision loss function to generate enhanced images that are perceptually similar across different exposure settings, while preserving spatial resolution and reducing artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are trained for standard or low exposure images, then image enhancement performance improves for those specific exposure conditions, but global uniformity across varying exposure inputs deteriorates
Solution Approach 1:
The patent transforms the input image from linear exposure space to logarithmic exposure space using the transformation log(I+1), where I is the input image intensity. This parameter transformation allows the neural network to learn exposure-invariant features, enabling the model to generalize across different exposure conditions while maintaining enhancement performance. The logarithmic transformation compresses the dynamic range and creates a more uniform representation that is less sensitive to exposure variations.
2Illumination intensity
If aggressive enhancement is applied to low exposure images, then visibility of dark regions improves, but artifacts such as halation and noise increase
Solution Approach 1:
The patent employs a feedback mechanism where the enhanced image is compared against the original input image, and the difference (error signal) is used to guide further refinement. The neural network processes this error signal to iteratively reduce artifacts while preserving enhancement benefits. This feedback loop allows the system to distinguish between genuine structural information and artifact noise, correcting halation and noise amplification while maintaining improved visibility in dark regions.
3Manufacturing precision
If high computational complexity is used to preserve high-frequency content, then detail preservation improves, but processing efficiency deteriorates
Solution Approach 1:
The patent applies preliminary logarithmic exposure transformation and normalization to the input image before feeding it to the neural network. This preprocessing step transforms the data into a more favorable representation space where high-frequency content is more easily preserved with reduced computational complexity. The transformation prepares the input in advance, allowing the network to focus computational resources on detail preservation rather than basic exposure correction.
Data Source
AI summary
A system for training a neural network includes a neural network configured to receive a training input in an image space and produce an enhanced image. The system further includes an error signal generator configured to compare the enhanced image to a ground truth and generate an error signal that is communicated back to the neural network to train the neural network. Additionally, the system includes a neural input enhancer configured to modify the training input in response to receiving at least one of an output from the neural network or the error signal. Modifying the training input improves one of an efficiency or a training result of the neural network beyond the communication of the error signal to only the neural network.