RAW Data Image Recognition via RGB Inversion

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing image recognition processing techniques require RGB data, which is generated by demosaicing RAW data, resulting in increased data volume and loss of texture information, limiting recognition accuracy.

Innovation Solution

An image processing apparatus and method that convert RGB data to RAW data, allowing image recognition processing to be performed directly on RAW data, thereby reducing resource usage and preserving texture information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If RGB data is used for image recognition processing, then recognition processing can be implemented, but data volume increases three times compared to RAW data

Engineering Contradiction:

Improvedata volumeVSAvoidtexture information

Core Design Contradiction:

Quantity of substanceVSLoss of information

Solution Approach 1:

Instead of converting RAW data to RGB data for recognition processing, the patent inverts the approach by converting RGB data back to RAW data format. This allows the recognition processing to be performed on RAW data, thereby reducing data volume while preserving texture information that would otherwise be lost in RGB conversion.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent changes the data format parameter from RGB to RAW, fundamentally altering how image data is represented. By using a conversion model trained to transform RGB data back to RAW format, the system maintains the benefits of RGB capture while achieving the efficiency of RAW data processing.

Inventive Principle:

Principle #35Parameter changes

2Loss of information

If RGB data is generated by demosaicing RAW data, then image recognition processing can be performed, but texture information is lost

Engineering Contradiction:

Improvetexture informationVSAvoidconversion process

Core Design Contradiction:

Loss of informationVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-training a conversion model to transform RGB data back to RAW data format. This pre-trained model is then used during inference to convert captured RGB data into RAW format before recognition processing, thereby preserving texture information without requiring complex real-time demosaicing adjustments.

Inventive Principle:

Principle #10Preliminary action

3Adaptability or versatility

If learning data in RGB format is converted to RAW format, then RAW data recognition can be trained, but conversion complexity increases

Engineering Contradiction:

Improvelearning data formatVSAvoidconversion process

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent uses copying by creating a learned mapping from RGB data to RAW data through training on paired datasets. Once this conversion model is trained, it can be repeatedly applied to convert any RGB data to RAW format without requiring complex real-time processing, as the conversion logic has been pre-established during the training phase.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS20250191233A1Image processing apparatus, image processing method, image conversion apparatus, image conversion method, ai network generation apparatus, ai network generation method, and program

Publication Date: 2025.06.12 SONY SEMICON SOLUTIONS CORP

AI summary

The present disclosure relates to an image processing apparatus, an image processing method, an image conversion apparatus, an image conversion method, an AI network generation apparatus, an AI network generation method, and a program that allow implementation of image recognition processing based on RAW data.Image recognition processing based on RAW data is implemented by generating a format conversion section that converts RGB data to RAW data by adversarial training and converting learning data including RGB data and a recognition result to learning data including RAW data and the recognition result to use the converted learning data for learning. The present disclosure can be applied to an image recognition apparatus.