Unpaired Document Image Enhancement via Latent Space Translation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current document image enhancement techniques face challenges such as the need for paired dirty and clean images for training, which raises privacy concerns and limits adaptability across different domains, and existing models are not effective in handling unknown degradation functions, especially in real-world applications.

Innovation Solution

The use of two separate implicit probabilistic models, where one translates electronic document images from a source space to a latent space and the other from the latent space to a target space, allowing for unpaired image translation and preserving data privacy by only accessing one set of images during training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If paired dirty and clean images are used for training, then model training effectiveness is improved, but data privacy is compromised

Engineering Contradiction:
Improvemodel training effectivenessVSAvoiddata privacy
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent introduces a latent space as an intermediary between the source domain (dirty images) and target domain (clean images). The first implicit probabilistic model translates dirty images to the latent space, and the second model translates from the latent space to clean images. This intermediary representation allows the models to learn the degradation function without directly accessing paired dirty-c clean image data, thus preserving data privacy while maintaining training effectiveness.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Device complexity

If a single model is used for image enhancement, then device complexity is reduced, but adaptability to different degradation types is worsened

Engineering Contradiction:
Improvemodel structure simplicityVSAvoidadaptability to different degradation types
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent divides the image enhancement task into two separate implicit probabilistic models: the first model handles translation from source domain to latent space, and the second model handles translation from latent space to target domain. This segmentation allows each model to specialize in specific aspects of the degradation function, improving adaptability to different degradation types while keeping the overall system manageable through the shared latent space framework.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If paired training data is required, then training data availability is improved, but applicability to real-world scenarios is worsened

Engineering Contradiction:
Improvetraining data availabilityVSAvoidapplicability to real-world scenarios
Core Design Contradiction:
Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The patent inverts the traditional supervised learning approach by training models on unpaired data. Instead of requiring dirty-c clean image pairs, the method trains the first model on dirty images to map to a latent space, and the second model on clean images to map from the latent space. This inversion allows the use of readily available unpaired training data while maintaining effectiveness in real-world scenarios where paired data is rarely available.

Inventive Principle:
Principle #13The other way round (Inversion)

Data Source

PatentEP4462382A1Model based document image enhancement
Publication Date: 2024.11.13 INTUIT INC
  • EP4462382A1 patent drawingFigure 1
  • EP4462382A1 patent drawingFigure 2
  • EP4462382A1 patent drawingFigure 3

AI summary

Systems and methods are disclosed for model based document image enhancement. Instead of requiring paired dirty and clean images for training a model to clean document images (which may cause privacy concerns), two models are trained on the unpaired images such that only the dirty images are accessed or only the clean images are accessed at one time. One model is a first implicit model to translate the dirty images from a source space to a latent space, and the other model is a second implicit model to translate the images from the latent space to clean images in a target space. The second implicit model is trained based on translating electronic document images in the target space to the latent space. In some implementations, the implicit models are diffusion models, such as denoising diffusion implicit models based on solving ordinary differential equations.