Neural Network Image Resolution Enhancement Using Secondary Camera Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image enhancement techniques using neural networks often introduce artifacts and distortions when scaling low-resolution images to higher resolutions, making them impractical for real-world applications, especially in portable devices with limited camera capabilities.

Innovation Solution

A method involving a neural network trained using image data from a secondary camera with lower resolution, scaled to match the resolution of a primary camera, incorporating error and distortion metrics to improve scaling accuracy and reduce artifacts, allowing for efficient image enhancement from low to high resolution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If neural networks are used to enhance image resolution by inferring missing pixels, then image resolution is improved, but artifacts and distortions are introduced into the enhanced image

Engineering Contradiction:
Improveimage resolutionVSAvoidartifacts and distortions
Core Design Contradiction:
Measurement precisionVSObject-generated harmful factors

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network using paired low-resolution and high-resolution image datasets before actual image enhancement. This pre-training phase allows the network to learn proper pixel inference patterns and avoid introducing artifacts during the actual enhancement process, as the network has already been conditioned on correct mappings between resolution levels.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback mechanisms through loss functions that compare the neural network's enhanced output against ground truth high-resolution images during training. This feedback loop continuously adjusts the network's weights and parameters to minimize errors and prevent artifact generation, ensuring the network learns to produce accurate results rather than spurious patterns.

Inventive Principle:
Principle #23Feedback

2Measurement precision

If neural networks are trained to improve image resolution, then image quality is enhanced, but training complexity and computational resources increase

Engineering Contradiction:
Improveimage qualityVSAvoidtraining complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the training process into distinct phases: pre-training on large paired datasets to learn general features, and fine-tuning on specific application scenarios. This segmented approach reduces overall training complexity by breaking down the monolithic training task into manageable stages with different computational requirements and objectives.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses preliminary action by pre-training the neural network architecture on large-scale paired image datasets before deployment. This pre-training establishes a strong foundation of learned features that reduces the computational burden during actual image enhancement operations, as the network has already acquired general image understanding and resolution-upscaling patterns.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If images are scaled from low to high resolution, then resolution is improved, but errors and distortions are introduced

Engineering Contradiction:
ImproveresolutionVSAvoidaccuracy of scaled image
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements feedback through loss functions that continuously compare the scaled output against ground truth high-resolution images during training. This feedback mechanism allows the neural network to learn the correct mappings between low and high resolution, minimizing errors and distortions by adjusting parameters based on the difference between predicted and actual high-resolution pixel values.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent replaces traditional mechanical interpolation methods with a neural network-based system that learns complex non-linear mappings. This substitution enables more accurate resolution enhancement by using data-driven patterns rather than simple mathematical interpolation, thereby reducing errors and maintaining higher reliability in the scaled images.

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

Data Source

PatentEP3631750B1Image resolution enhancement
Publication Date: 2022.10.12 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3631750B1 patent drawingFigure 1
  • EP3631750B1 patent drawingFigure 2A
  • EP3631750B1 patent drawingFigure 2B

AI summary

Resolution enhancement techniques are described. An apparatus may receive first image data at a first resolution, and second image data at a resolution less than the first resolution. The second image data may be scaled to the first resolution and compared to the first image data. Application of a neural network may scale the first image data to a resolution higher than the first resolution. The application of the neural network may incorporate signals based on the scaled second image data. The signals may include information obtained by comparing the scaled second image data to the resolution of the first image data.