Unsupervised Neural Network HDR Image Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing HDR image compression techniques introduce unwanted artifacts and require manual tuning of calibration parameters, making them inefficient and impractical for widespread use.
Innovation Solution
An unsupervised neural network training methodology that uses a loss function to preserve contrast and compress HDR images, eliminating the need for ground truth images and manual parameter tuning, and employs iterative backpropagation to minimize the loss function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing analytical approaches are used for HDR compression, then compression is achieved, but unwanted image artifacts are introduced
Solution Approach 1:
The patent replaces traditional analytical/mathematical compression approaches with a neural network-based system. The neural network learns optimal compression mappings from training data, substituting rigid mathematical transformations with adaptive learned transformations that preserve image quality while achieving compression.
Solution Approach 2:
The patent changes the fundamental parameters of the compression system by using learned weights and biases in a neural network instead of fixed analytical parameters. The system transforms input HDR images through multiple layers with adjustable parameters that are optimized during training to minimize artifacts while maintaining compression efficiency.
2Object-affected harmful factors
If supervised neural network training is used for HDR compression, then image quality can be improved, but ground truth images and manual parameter tuning are required
Solution Approach 1:
The patent implements self-service training where the neural network trains itself using only the input HDR images without requiring separate ground truth images or manual calibration. The network learns to compress and then reconstruct images by minimizing reconstruction error, automatically adapting to preserve image quality without external supervision or manual parameter tuning.
Solution Approach 2:
Instead of the traditional supervised approach where ground truth images guide the training, the patent inverts the methodology by using only input images and having the network learn through self-reconstruction. The training process is reversed from supervised learning to unsupervised self-learning, eliminating the need for expensive ground truth data.
3Manufacturing precision
If manual tuning of calibration parameters is performed, then compression accuracy can be improved, but time consumption and practicality are reduced
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network on a large dataset of HDR images before deployment. During this offline training phase, the network learns optimal compression parameters and mappings, so that when deployed, it can immediately compress images with high accuracy without requiring any manual parameter tuning or calibration for each new image.
Data Source
AI summary
Techniques are provided for unsupervised training of a neural network to perform compression of a high dynamic range (HDR) image. A methodology implementing the techniques according to an embodiment includes performing global tone mapping on an HDR training image to generate a low dynamic range (LDR) training image. The method also includes applying the neural network to the HDR training image and the LDR training image to generate a delta image representing image detail lost in the global tone mapping operation. The method further includes summing the delta image with the LDR training image to generate an output training image, and generating a loss function calculated from a weighted sum of a contrast loss and a compression loss. The contrast loss is based on the output training image and the HDR training image, and the compression loss is based on the output training image and the LDR training image.