Image Classifier Sensitivity Measurement via Intermediate Image Generation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Mass-produced product quality control using image classifiers faces challenges in accurately distinguishing between 'OK' and 'not OK' products due to sensitivity issues related to changes in input images, such as noise and lighting conditions, leading to incorrect classifications and potential rejection of flawless products or failure to detect defects.

Innovation Solution

A method involving the creation of intermediate images with reduced information content or poorer signal-to-noise ratios, generated by trained generators, to increase classification uncertainty, allowing for small interventions to change class assignments while maintaining realistic images, and using these modifications to evaluate and improve the sensitivity of image classifiers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the image classifier is trained to be robust against changes in input image, then the reliability of classification is improved, but the sensitivity to detect actual defects deteriorates

Engineering Contradiction:
Improveclassification reliabilityVSAvoiddefect detection precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies partial action by introducing intermediate images with partial degradation (reduced information content or added noise) rather than complete transformation. This allows the classifier to learn robustness to specific types of perturbations while maintaining sensitivity to actual defects. The degradation is controlled and partial, enabling the system to distinguish between meaningful defects and benign variations.

Inventive Principle:
Principle #16Partial or excessive action

2Reliability

If the information content of the input image is reduced to increase robustness, then the reliability against noise is improved, but the measurement precision of defect detection worsens

Engineering Contradiction:
Improverobustness against noiseVSAvoiddefect detection precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by pre-processing input images to create intermediate representations with reduced information content or added noise before classification. This preliminary transformation allows the classifier to learn invariant features that are robust to noise and lighting variations, while still preserving the essential defect characteristics needed for accurate detection.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If the image classifier is made highly sensitive to changes, then the detection precision is improved, but the reliability increases false rejections worsens

Engineering Contradiction:
Improvedefect detection precisionVSAvoidfalse rejection rate
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces intermediate images as mediators between the original input and the classifier decision. These intermediate representations serve as a buffer that filters out benign variations while preserving defect signals. The classifier operates on these stabilized intermediate representations, which reduces false rejections while maintaining detection precision through the selective preservation of important features.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3923193B1Measurement of sensitivity of image classifiers against changes in the input image
Publication Date: 2024.03.27 ROBERT BOSCH GMBH
  • EP3923193B1 patent drawingFigure 1
  • EP3923193B1 patent drawingFigure 2
  • EP3923193B1 patent drawingFigure 3

AI summary

Method (100) for measuring the sensitivity (2*) of an image classifier (2), which assigns an input image (1) to one or more classes (3a-3c) of a given classification, against changes to the input image (1), comprising the steps: • the input image (1) is mapped by at least one given operator (4) to an intermediate image (5) (110) which has a lower information content and/or a worse signal-to-noise ratio compared to the input image (1); • at least one generator (6) is provided (120) which is trained to produce realistic images that are assigned by the image classifier (2) to a specific class (3a-3c) of the given classification; with this generator (6) a modification (7) of the input image (1) is generated from the intermediate image (5) (130).