Image Domain Membership via Predefined Attribute Catalogs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for specifying the operational design domain (ODD) of machine learning models for safety-critical applications like automated driving are primarily manual and inefficient, lacking a systematic approach to ensure the domain and distribution of training examples match real-world scenarios.
Innovation Solution
A method using a predefined catalog of attributes to determine the domain and distribution of images, leveraging machine learning models to analyze and categorize images based on these attributes, enabling automatic creation and refinement of ODDs through generative models and textual information, and training additional models for specific tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a manual process is used to create the operational design domain (ODD) specification, then the domain definition can be customized and detailed, but the process is time-consuming and inefficient
Solution Approach 1:
The patent replaces the manual mechanical process of ODD specification creation with an automated machine learning-based system. The ML model automatically extracts domain characteristics from training images and generates ODD specifications, eliminating the need for manual creation while maintaining or improving accuracy through systematic analysis of actual training data distributions.
Solution Approach 2:
The system enables the ODD specification to be self-generated from the training data itself. By analyzing the attributes and distributions present in the training images, the system automatically creates an accurate domain definition without requiring external manual intervention, making the process self-service and highly efficient.
2Adaptability or versatility
If the operational design domain specification is detailed and comprehensive, then the model can function flawlessly across various conditions, but the complexity of specifying and managing the ODD increases
Solution Approach 1:
The patent segments the complex ODD specification into distinct attribute categories (e.g., environmental conditions, infrastructure, traffic conditions). Each category is independently analyzed and specified based on the training data, making the overall complex specification manageable through modular organization of domain characteristics.
Solution Approach 2:
The system automatically determines domain parameters and their ranges by analyzing the actual distribution of attributes in the training data. This data-driven approach to parameter specification ensures comprehensive coverage of operating conditions while simplifying the specification process through automated parameter extraction rather than manual definition.
3Reliability
If machine learning models are trained on large numbers of training examples to improve generalization, then the model performance improves, but ensuring the training domain matches real-world scenarios becomes more difficult
Solution Approach 1:
The patent implements a feedback mechanism where the ODD specification is continuously refined based on analyzing the actual domain characteristics of training images. The system compares the extracted training domain attributes against the ODD specification and automatically adjusts the specification to ensure alignment, creating a closed-loop system that guarantees domain matching as training data is added.
Data Source
Figure 1
Figure 2
AI summary
Method (100) for determining a domain and/or distribution (2) to which an image (1) belongs, comprising the steps: • a predetermined catalogue (4) of attributes (5) is provided (110), with respect to which an image (1) can assume different values (6), wherein ∘ this catalogue (4) of attributes (5) relates to a higher-level domain and/or distribution (3) of images (1) and ∘ combinations of values (6) of the attributes (5) indicate that the image belongs to sub-domains or sub-distributions (2) of the higher-level domain orDisplay distribution (3); • the image (1) is fed to one or more trained machine learning models (7) as input (120); • from the outputs (7a) then supplied by the one or more trained machine learning models (7), the values (6) of the attributes (5) from the catalog (4) for the image (1) are determined (130); and • based on these values (6) of the attributes (5), the affiliation of the image (1) to a sub-domain or sub-distribution (2) of the parent domain and/or distribution (3) is determined (140).