RGB-to-NIR Image Training With Round-Trip Feedback Segmentation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing techniques for generating near-infrared (NIR) facial images using deep learning models are influenced by the quality of input RGB images, often resulting in unusable NIR images due to significant errors.

Innovation Solution

A method involving a first generative model to convert RGB images to NIR images, a second generative model to convert NIR images back to RGB, and a segmentation model to enhance object distinctiveness, with parameter optimization based on errors between input and output images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If existing deep learning models are used to generate NIR images from RGB images, then the generation process can be automated, but the quality of NIR images deteriorates due to significant errors influenced by input RGB image quality

Engineering Contradiction:

Improveautomation of NIR image generationVSAvoidquality of generated NIR images

Core Design Contradiction:

Extent of automationVSManufacturing precision

Solution Approach 1:

The patent introduces a feedback mechanism where the generated NIR image is converted back to RGB space using a second deep learning model, and the converted image is compared with the original RGB image. The error between these images is calculated and used to iteratively adjust the parameters of the first deep learning model, thereby improving the quality of generated NIR images while maintaining automation.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent uses a second deep learning model as an intermediary to convert the generated NIR image back to RGB space. This intermediary enables the comparison and error calculation process, which is essential for optimizing the first model's parameters without requiring direct access to ground truth NIR images.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If parameters of the first deep learning model are set to minimize error between first RGB image and second RGB image, then the round-trip conversion accuracy is improved, but the quality of intermediate NIR images deteriorates

Engineering Contradiction:

Improveaccuracy of RGB-NIR-RGB round-trip conversionVSAvoidquality of intermediate NIR images

Core Design Contradiction:

Measurement precisionVSManufacturing precision

Solution Approach 1:

The patent employs feedback by calculating the error between the original RGB image and the converted RGB image (obtained from NIR via second model), and using this error to iteratively optimize the parameters of the first deep learning model. This feedback loop ensures that minimizing round-trip conversion error simultaneously improves intermediate NIR image quality.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent performs preliminary optimization of the first deep learning model parameters by using the feedback mechanism before final NIR image generation. This preliminary action ensures that the model is pre-adjusted to produce high-quality NIR images that maintain accuracy through the round-trip conversion process.

Inventive Principle:

Principle #10Preliminary action

3Device complexity

If a single deep learning model is used for RGB to NIR conversion, then the system complexity is reduced, but the reliability of generated NIR images deteriorates due to lack of error correction mechanism

Engineering Contradiction:

Improvenumber of deep learning modelsVSAvoidreliability of generated NIR images

Core Design Contradiction:

Device complexityVSReliability

Solution Approach 1:

The patent introduces a feedback mechanism using a second deep learning model to convert NIR images back to RGB space. The error between the original and converted RGB images is calculated and used to iteratively optimize the first model's parameters, significantly improving the reliability of generated NIR images.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent combines multiple deep learning models (first model for RGB-NIR conversion, second model for NIR-RGB conversion, and segmentation model) into an integrated system. This merging creates a comprehensive error correction mechanism that enhances reliability while managing system complexity through unified parameter optimization.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS12567149B2Method and apparatus for training image processing model, and storage medium storing instructions to perform method for training image processing model

Publication Date: 2026.03.03 SUPREMA INC

AI summary

There is provided a method for training an image processing model. The method comprises preparing the image processing model including a first generative model, a second generative model, and a segmentation model; inputting a first RGB image into the first generative model and determining a near-infrared ray (NIR) image as an output of the first generative model; inputting the NIR image into the second generative model and determining a second RGB image as an output of the second generative model; inputting the first RGB image and the second RGB image into a segmentation model and determining a first segmentation image and a second segmentation image as outputs of the segmentation model; and setting parameters of the first generative model based on an error between the first RGB image and the second RGB image and an error between the first segmentation image and the second segmentation image.