Image Generator Network Blur Adaptation for Super-Resolution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Classical interpolation methods in smartphones result in blurry, low perceptual quality images when attempting high digital zoom, and existing machine learning-based super-resolution solutions require large datasets and deep knowledge of the device's optics to simulate realistic blur characteristics.

Innovation Solution

A method for training an image generator network using a discriminator network to adapt the blur module, allowing the generation of simulated images that mimic the blur characteristics of a different device without requiring knowledge of either device's specifics, by iteratively refining the network to minimize differences between low-resolution and simulated images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If classical interpolation methods are used for digital zoom, then device complexity is reduced, but image quality deteriorates (blurry results)

Engineering Contradiction:
Improveimage processing complexityVSAvoidimage quality
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent replaces classical mechanical interpolation algorithms with a neural network-based super-resolution system. The neural network learns optimal pixel reconstruction patterns from training data, substituting traditional mathematical interpolation with learned patterns that preserve image quality while enabling digital zoom capabilities.

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

Solution Approach 2:

The patent performs preliminary training of the neural network with paired high-resolution and low-resolution images before actual image processing. This pre-training phase establishes the network's ability to reconstruct high-quality images from low-resolution inputs, so that during operation, only the trained network needs to be applied without requiring complex real-time calculations.

Inventive Principle:
Principle #10Preliminary action

2Manufacturing precision

If machine learning-based super-resolution is implemented, then image quality is improved, but data set requirements increase

Engineering Contradiction:
Improveimage qualityVSAvoiddata set size
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

Instead of directly training a super-resolution network with large datasets of high-resolution images, the patent inverts the approach by first training a network to generate low-resolution versions of high-resolution images, then using this generated data to train the final super-resolution network. This inversion reduces the need for large amounts of annotated high-resolution training data.

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

Solution Approach 2:

The patent uses the high-resolution images to generate synthetic low-resolution training pairs through the trained downscaling network. These copied and transformed images serve as training data, reducing the need for collecting and annotating large numbers of real low-resolution reference images.

Inventive Principle:
Principle #26Copying

3Measurement precision

If device-specific blur knowledge is incorporated, then simulation accuracy is improved, but ease of operation deteriorates

Engineering Contradiction:
Improveblur simulation accuracyVSAvoidtraining simplicity
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The patent enables the neural network to automatically learn device-specific blur characteristics directly from image pairs without requiring manual input of optical parameters. The network self-adjusts its blur simulation by comparing generated low-resolution images with actual device outputs, eliminating the need for users to provide detailed device specifications.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent allows the blur characteristics to be learned as trainable parameters within the neural network rather than fixed values based on device specifications. This enables the system to adapt blur parameters dynamically during training, automatically adjusting to match the specific device's optical characteristics without manual configuration.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4398184A1Image generator network and device
Publication Date: 2024.07.10 BEIJING XIAOMI MOBILE SOFTWARE CO LTD
  • EP4398184A1 patent drawingFigure 1
  • EP4398184A1 patent drawingFigure 2
  • EP4398184A1 patent drawingFigure 3

AI summary

Method for training an image generator network, including the steps of: Acquiring a high resolution (HR) image of a scene by a first device; Acquiring a low resolution (LR) image of the same scene by a second device, wherein the LR image has a lower resolution as the HR image; Providing the HR image to the image generator network to generate a simulated image, wherein the image generator network includes a downscale module to downscale the HR image and a blur module to adapt the blur of the simulated image; Providing the LR image and the simulated image to a discriminator network to judge a difference between the simulated image and the LR image; Adapting the image generator network for training based on the result of the discriminator network to minimize the difference between the simulated image and the LR image.