Permutation-Invariant HDR Imaging via Symmetric Feature Pooling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for high dynamic range (HDR) imaging face challenges in capturing detailed scenes with broad luminosity ranges, particularly in single-image HDR processing, which relies on hallucinating details that are under- or over-exposed, and multi-frame HDR methods struggle with accurate alignment and motion issues.

Innovation Solution

A permutation invariant deep learning model that performs symmetric pooling operations on feature representations from one or multiple low dynamic range (LDR) images, enabling denoising, dequantization, and reconstruction of under- and over-exposed regions, allowing for flexible input ordering and handling of images from different sensors with varying noise levels.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Illumination intensity

If multi-frame HDR fusion is used to capture broad luminosity ranges, then dynamic range coverage is improved, but alignment accuracy and processing reliability deteriorate due to motion and registration difficulties

Engineering Contradiction:
Improvedynamic range coverageVSAvoidalignment accuracy
Core Design Contradiction:
Illumination intensityVSReliability

Solution Approach 1:

The patent segments the multi-frame HDR problem into independent single-image HDR processing units. Each input image is processed separately through the neural network to generate feature representations, which are then combined through symmetric pooling. This segmentation eliminates the need for frame alignment while maintaining the benefit of multiple exposure values, resolving the contradiction between dynamic range coverage and alignment reliability.

Inventive Principle:
Principle #1Segmentation

2Reliability

If single-image HDR processing is used to avoid alignment issues, then processing reliability is improved, but luminosity range coverage deteriorates due to reliance on hallucinating details

Engineering Contradiction:
Improveprocessing reliabilityVSAvoidluminosity range coverage
Core Design Contradiction:
ReliabilityVSIllumination intensity

Solution Approach 1:

The patent merges multiple single-image HDR processing results through symmetric pooling operations. The neural network processes each input image independently to maintain reliability, then the feature representations from multiple images are combined using symmetric pooling (max pooling, average pooling, or sum pooling). This merging allows the system to achieve broad luminosity range coverage by leveraging information from multiple exposures while maintaining the reliability of single-image processing.

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If permutation invariant processing is implemented to handle flexible input ordering, then adaptability is improved, but computational complexity increases due to symmetric pooling operations

Engineering Contradiction:
Improveinput ordering flexibilityVSAvoidcomputational complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements universal symmetric pooling operations that can process any number of input images in any order without requiring reconfiguration. The same pooling mechanism (max, average, or sum) works regardless of input count or sequence, providing permutation invariance. This universality achieves adaptability to flexible input ordering while keeping computational complexity manageable through efficient pooling implementations that scale well with the number of inputs.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4222688B1Permutation invariant high dynamic range imaging
Publication Date: 2025.07.23 HUAWEI TECH CO LTD
  • EP4222688B1 patent drawingFigure 1(a)~2(b)
  • EP4222688B1 patent drawingFigure 3
  • EP4222688B1 patent drawingFigure 4(a)~4(b)

AI summary

Described herein is an image processing apparatus (1200) for forming an enhanced image, the apparatus comprising one or more processors (1204) configured to: receive (1101) one or more input images (301, 302, 404, 405, 406, 501, 502, 503, 601, 602, 603); form (1102) from each of the one or more input images a respective feature representation, each feature representation representing features of the respective input image; and subject (1103) the one or more feature representations to a symmetric pooling operation (305) to form an enhanced image (408, 505, 608) from at least some of the features of the one or more feature representations identified by the symmetric pooling operation. The apparatus may result in images with increased photoreceptive dynamic range, increased bit depth and signal-to-noise ratio, with less quantization error and richer colour representation.