Image Resizing Aligned With ML Inference Requirements

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The lack of knowledge about downstream consumer requirements in image resizing algorithms leads to poor inference quality in machine learning models, as resizing algorithms optimized for visual quality may not align with the specific needs of machine learning inference tasks.

Innovation Solution

A resizer is made aware of downstream consumer information, adapting its resizing algorithm based on metadata or profiling techniques to match the requirements of the downstream consumer, optimizing for inference quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If a resizing algorithm optimized for visual quality (e.g., bilinear) is used, then visual quality is improved, but inference quality of machine learning models deteriorates

Engineering Contradiction:
Improvevisual qualityVSAvoidinference quality
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The resizer dynamically adapts its resizing algorithm based on downstream consumer information. Instead of using a fixed algorithm optimized for visual quality, the system selects different resizing algorithms (e.g., nearest neighbor, bilinear, bicubic, lanczos) depending on the specific requirements of the downstream consumer, such as machine learning inference tasks that benefit from sharper edges and more defined features

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes the parameters of the resizing operation by selecting different algorithms based on consumer intent metadata. The downstream consumer information includes identifiers for preferred resizing algorithms and parameters, allowing the resizer to adjust its behavior to match the specific needs of machine learning models versus other consumers like playback applications

Inventive Principle:
Principle #35Parameter changes

2Ease of operation

If a generic resizing algorithm is used without knowledge of downstream consumer requirements, then ease of operation is improved, but inference quality deteriorates

Engineering Contradiction:
Improveease of useVSAvoidinference quality
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The downstream consumer (e.g., machine learning model) provides self-service by embedding or signaling its own requirements upstream to the resizer. The consumer intent metadata is generated by the downstream consumer itself, allowing it to specify its preferred resizing algorithm and parameters, eliminating the need for manual configuration or generic defaults

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system implements feedback by having the downstream consumer communicate its requirements back to the resizer through consumer intent metadata. This feedback loop allows the resizer to adjust its algorithm selection based on the actual needs of the downstream consumer, improving inference quality while maintaining ease of operation through automated adaptation

Inventive Principle:
Principle #23Feedback

3Adaptability or versatility

If different resizing algorithms are used for different downstream consumers, then adaptability is improved, but device complexity increases

Engineering Contradiction:
Improveadaptability to consumer requirementsVSAvoidresizer complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The resizer is designed as a universal component that can serve multiple downstream consumers with different requirements. By implementing a selection mechanism that chooses from multiple resizing algorithms based on consumer intent metadata, the single resizer component achieves multi-functionality, adapting to machine learning inference, playback, and other consumers without requiring separate dedicated resizers for each

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4712016A1Resizing for enhanced inference
Publication Date: 2026.03.18 INTEL CORP
  • EP4712016A1 patent drawingFigure 1
  • EP4712016A1 patent drawingFigure 2
  • EP4712016A1 patent drawingFigure 3

AI summary

The lack of knowledge about a downstream consumer using a resized image can lead to poor inference quality of a machine learning model. Inference quality can be improved when the resizing algorithm to produce resized images closely matches the one used during training of the machine learning model. To achieve this technical task, a resizer can be made aware of downstream consumer information and apply a suitable resizing algorithm. In one scenario, the downstream consumer information is received as metadata from a downstream process. In another scenario, an optimal resizing option can be determined to maximize inference quality. In yet another scenario, a likely resizing option can be determined by assessing a filtering profile determined based on a known original image and a known resized image.