Depth Refocusing with Differentiable Masks for Learnable Imaging

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing methods for computationally refocusing digital images are not differentiable, preventing their integration into learnable pipelines and limiting the ability to adapt and improve the refocusing process through machine learning.

Innovation Solution

Implement a differentiable refocusing algorithm using a differentiable function to generate depth masks, allowing the refocusing process to be part of a learnable pipeline, and enabling the use of machine learning techniques to adapt and improve the image processing model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional non-differentiable refocusing algorithms are used, then refocusing functionality is achieved, but integration into learnable pipelines is prevented

Engineering Contradiction:
Improveintegration into learnable pipelinesVSAvoidalgorithm differentiability requirements
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent replaces traditional non-differentiable image refocusing operations with a differentiable neural network-based refocusing algorithm. This substitution enables the refocusing process to be integrated into learnable deep learning pipelines, allowing end-to-end training and optimization while maintaining the core functionality of generating refocused images from input image sequences.

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

2Manufacturing precision

If machine learning techniques are applied to improve refocusing, then image quality is enhanced, but computational complexity increases

Engineering Contradiction:
Improveimage refocusing qualityVSAvoidcomputational resources required
Core Design Contradiction:
Manufacturing precisionVSPower

Solution Approach 1:

The patent employs a dynamic and adaptive refocusing algorithm that adjusts its computational complexity based on the specific input characteristics. The neural network-based approach learns optimal refocusing parameters from training data, enabling efficient processing while maintaining high image quality. The system dynamically adapts to different scenes and focusing requirements without requiring excessive computational resources for every operation.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3931795B1Depth of field image refocusing
Publication Date: 2025.11.26 HUAWEI TECH CO LTD
  • EP3931795B1 patent drawingFigure 1
  • EP3931795B1 patent drawingFigure 2
  • EP3931795B1 patent drawingFigure 3

AI summary

An image processing device comprising a processor configured to generate a refocused image from an input image and a map indicating depth information for the image, by the steps of: for each of a plurality of planes associated with respective depths within the image: generating a depth mask having values indicating whether regions of the input image are within a specified range of the plane, wherein an assessment of whether a region is within the specified range of the plane is made through the evaluation of a differentiable function of the range between regions of the input image and the plane as determined from the map; generating a masked image from the input image and the generated depth mask; refocusing the masked image using a blurring kernel to generate a refocused partial image; and generating the refocussed image from the plurality of refocussed partial images.