Image Upscaling Color-Space Loss Balancing for Accurate Color Reproduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional ML-based image super-resolution upscaling processes suffer from color shifts and loss of clear color separation among pixels, particularly when using an RGB image format without a deep neural network architecture, leading to issues like human skin tone distortion and objects losing clear color separation.

Innovation Solution

The process involves converting RGB-formatted data to color-space-separable-formatted data, such as YCbCr or Lab, using color components like CbCr channels as additional input, and adjusting the balance between RGB and color loss functions by controlling their weights to prioritize overall image quality and color reproduction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If RGB image format is used for input without deep neural network architecture, then the upscaling process is simpler and faster, but color shifts and loss of clear color separation occur

Engineering Contradiction:
Improveupsaling process speedVSAvoidcolor reproduction accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent segments the color reproduction task by introducing a dedicated color loss function that separately evaluates color accuracy from the overall image quality. This is achieved by converting predicted images to YCbCr color space and computing color differences in the CbCr channels, allowing independent optimization of color reproduction without requiring complex deep neural network architectures.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter space for color evaluation by transitioning from RGB color space to YCbCr color space. This parameter transformation enables separate control and evaluation of luminance (Y) and color (CbCr) components, allowing the loss function to specifically target color accuracy while maintaining overall image quality through balanced weighting.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If conventional ML-based upscaling is used, then processing is faster, but color shifts and distortion of human skin tone occur

Engineering Contradiction:
Improveprocessing speedVSAvoidcolor accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent implements feedback control by introducing a color loss function that provides explicit feedback on color reproduction accuracy during the training process. The color loss is computed by comparing CbCr channels of original and predicted images, and this feedback is used to adjust model parameters to minimize color errors, ensuring reliable color reproduction in the final upscaling model.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent performs preliminary action by pre-defining the color loss function and color space transformation (RGB to YCbCr) before the actual upscaling training. This preliminary setup establishes the color accuracy evaluation framework in advance, allowing the model to learn color-preserving transformations from the outset rather than requiring post-processing corrections.

Inventive Principle:
Principle #10Preliminary action

3Device complexity

If RGB-formatted data is used without color-space separation, then the data processing is simpler, but clear color separation among pixels is lost

Engineering Contradiction:
Improvedata processing complexityVSAvoidcolor separation information
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent segments the image data into luminance and color components by transforming from RGB to YCbCr color space. This segmentation separates the Y (luminance) channel from the CbCr (color) channels, allowing the loss function to specifically evaluate and preserve color separation information in the CbCr channels while maintaining overall processing efficiency.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces YCbCr color space as an intermediary representation between RGB input and the loss computation. This intermediary transformation enables explicit color separation evaluation without requiring complex processing of raw RGB data, as the CbCr channels directly represent color information that can be independently analyzed and optimized.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3844682B1Enhanced color reproduction for upscaling
Publication Date: 2025.12.24 SONY GROUP CORP
  • EP3844682B1 patent drawingFigure 1
  • EP3844682B1 patent drawingFigure 2
  • EP3844682B1 patent drawingFigure 3

AI summary

A method for enhancing color reproduction in an upscaling process, including: converting RGB-formatted data to color-space-separable-formatted data; sending the RGB-formatted data to a neural network for training to generate RGB-predicted data; converting the RGB-predicted data to color-space-separable-predicted data; computing a first loss function by calculating a first difference between the RGB-formatted data and the RGB-predicted data; extracting color-formatted data from the color-space-separable-formatted data; extracting color-predicted data from the color-space-separable-predicted data; computing a second loss function by calculating a second difference between the color-formatted data and the color-predicted data; and adjusting a balance between the first and second loss functions.