Single-Image HDR Generation via CNN Intensity Transformation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for generating high dynamic range (HDR) images require multiple images with different exposures, making the process time-consuming and complex, and struggle with color consistency issues due to varying illumination and camera settings, especially in applications like HDR imaging where input Low Dynamic Range (LDR) images need alignment and motion compensation.

Innovation Solution

An image processing device equipped with an intensity transformer using a convolutional neural network (CNN) that learns exposure transformation from source and target images, allowing it to generate an HDR image from a single input image by determining pixel intensities based on neighboring pixels, without relying on geometric correspondences or statistical properties, thus enabling exposure conversion and color adjustment.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Illumination intensity

If multiple images with different exposures are acquired for HDR imaging, then the dynamic range is increased, but the image acquisition time and complexity increase

Engineering Contradiction:
Improvedynamic rangeVSAvoidimage acquisition time
Core Design Contradiction:
Illumination intensityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training a neural network model on pairs of images with different exposures before actual HDR processing. This pre-learning phase enables the model to understand exposure transformations, allowing single-image HDR generation without requiring multiple exposure acquisitions during actual use, thus reducing time loss while maintaining dynamic range enhancement capability

Inventive Principle:
Principle #10Preliminary action

2Illumination intensity

If multiple images with different exposures are used for HDR imaging, then the dynamic range is increased, but the process complexity increases

Engineering Contradiction:
Improvedynamic rangeVSAvoidprocess complexity
Core Design Contradiction:
Illumination intensityVSDevice complexity

Solution Approach 1:

The patent replaces the mechanical/optical system of acquiring multiple physical images with different exposures with a computational system. A neural network model processes a single input image to generate multiple virtual images with different exposure levels, substituting the need for multiple camera exposures and complex alignment procedures with automated deep learning-based image transformation

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent creates virtual copies of the input image at different exposure levels through neural network processing. Instead of capturing multiple physical copies with different exposures, the model generates synthesized versions that mimic the appearance of differently exposed images, reducing the need for multiple acquisitions and subsequent complex merging operations

Inventive Principle:
Principle #26Copying

3Stability of the object's composition

If exposure conversion is performed on images with motion differences, then color consistency is improved, but the requirement for motion compensation and alignment increases complexity

Engineering Contradiction:
Improvecolor consistencyVSAvoidmotion compensation complexity
Core Design Contradiction:
Stability of the object's compositionVSDevice complexity

Solution Approach 1:

The patent replaces the complex mechanical/optical process of motion compensation and geometric alignment with a neural network-based approach. The model learns to handle motion differences between images by processing pixel intensities directly, eliminating the need for separate motion estimation, feature matching, and geometric transformation steps that would otherwise be required to achieve color consistency

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP3350767B1Exposure-related intensity transformation
Publication Date: 2020.06.03 HUAWEI TECH CO LTD
  • EP3350767B1 patent drawingFigure 1~2
  • EP3350767B1 patent drawingFigure 3~4
  • EP3350767B1 patent drawingFigure 5

AI summary

An image processing device for generating from an input image, which is a first image of a scene, an output image, which is a second image of the scene, wherein the output image corresponds to an exposure that is different from an exposure of the input image, the image processing device comprising an intensity transformer that is configured to determine one or more intensities of an output pixel, which is a pixel of the output image, based on an input patch around an input pixel, which is a pixel of the input image that corresponds to the output pixel.