Image Classifier Section Detection via Targeted Latent-Space Perturbation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image classifiers in quality control and automated vehicle driving lack effective methods to explain their decision-making processes, particularly in identifying decision-relevant features, leading to potential misclassifications with varying severity and difficulty in optimizing their performance.

Innovation Solution

A method involving convolutional layers to generate an intermediate product, applying perturbations based on counter images to determine decision-relevant parts through binary masks, and using quality functions to quantify the importance of these parts, allowing for model-independent analysis and optimization of image classifiers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If image classifiers are used for automated quality control and vehicle driving, then productivity and automation are improved, but the lack of explainability in decision-making processes leads to reduced reliability and increased difficulty in detecting misclassifications

Engineering Contradiction:
Improveautomated quality control and vehicle drivingVSAvoidtrustworthiness of classifier decisions
Core Design Contradiction:
Extent of automationVSReliability

Solution Approach 1:

The patent introduces an intermediary explanation system that mediates between the image classifier and the user. This system generates visual explanations (heatmaps, saliency maps) that act as a bridge, making the classifier's decision-making process transparent and interpretable without altering the classifier itself.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements feedback mechanisms where the explanation system provides information back to users about why certain decisions were made. This feedback loop allows users to understand classifier behavior, verify correctness, and improve trust in automated systems.

Inventive Principle:
Principle #23Feedback

2Device complexity

If existing image classifiers are deployed without explanation methods, then device complexity is reduced, but the inability to identify decision-relevant features leads to reduced measurement precision and increased loss of information

Engineering Contradiction:
Improvesimplicity of classifier deploymentVSAvoididentification of decision-relevant features
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The explanation system serves as an intermediary layer that processes the classifier's internal representations and translates them into interpretable visual formats, enabling precise identification of decision-relevant features without modifying the original classifier architecture.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms the high-dimensional internal representations of the classifier into two-dimensional visual explanations that can be directly interpreted by humans, adding a new dimension of interpretability without losing information about decision-relevant features.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Manufacturing precision

If more convolutional layers are used to improve classification accuracy, then manufacturing precision and reliability are improved, but the complexity of the latent space increases, making it harder to identify and correct misclassifications

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomplexity of latent space analysis
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The explanation system acts as an intermediary that simplifies the complex latent space by projecting it into visual representations that highlight only the most relevant features, making deep network decisions interpretable regardless of network depth.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent extracts the essential decision-relevant information from the complex latent space of deep convolutional networks, separating the crucial features from the redundant or noisy information, thereby simplifying analysis while maintaining accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

4Loss of information

If targeted modification in latent space is implemented, then loss of information is reduced and measurement precision is improved, but device complexity and computational requirements increase

Engineering Contradiction:
Improvepreservation of decision-relevant informationVSAvoidcomputational complexity of explanation system
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The explanation system extracts only the essential decision-relevant information from the latent space, discarding redundant data, thereby preserving critical information while minimizing computational overhead.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial modification to the latent space, focusing computational resources on modifying only the portions of the representation that are most relevant to the decision, rather than processing the entire latent space uniformly.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4012611B1Determination of image sections relevant for an image classifier by targeted modification in the latent space
Publication Date: 2025.08.20 ROBERT BOSCH GMBH
  • EP4012611B1 patent drawingFigure 1a
  • EP4012611B1 patent drawingFigure 1b
  • EP4012611B1 patent drawingFigure 2a~2d

AI summary

Method (100) for measuring the proportions (2a) of an input image (2) on which an image classifier (1) bases its decision on assigning this input image (2) to one or more classes of a given classification, comprising the steps: • the input image (2) is processed by the image classifier (1) through one or more convolution layers to form an intermediate (3) (110); • the intermediate (3) is mapped by the image classifier (1) to a classification score (7) with respect to at least one target class (120); • from one or more counter-images (5a-5c) that are preferentially assigned by the image classifier (1) to at least one class other than the target class, a perturbation (6) in the space of the intermediates (3) is determined (130); • at least one binary mask (4) is provided (140) which has the same number of pixels as the intermediate (3);• From the intermediate product (3), at least one modification (3*) is created (150) in which pixels defined by the binary mask (4) are replaced by corresponding pixels of the disturbance (6); • the modification (3*) is mapped by the image classifier (1) to a classification score (7*) with respect to a predefined class (160); • from this classification score (7*), a power function (8) is used to determine (170) the extent (8a) to which the binary mask (4) specifies the desired decision-relevant components (2a) of the input image (2).