ANN Plausibility Check via Spatial Relevance Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The reliability of object recognition in semi-automated driving systems using neural networks is compromised due to the lack of a robust plausibility check for the outputs of artificial neural networks (ANNs), which can lead to incorrect classification of traffic-relevant objects, especially in complex scenarios where multiple objects are present.
Innovation Solution
A method is introduced to assess the plausibility of ANN outputs by determining feature parameters for images, applying a relevance assessment function to identify the spatial contributions to classifications, and evaluating the applicability of these assessments using setpoint relevance assessments and quality criteria, ensuring invariance and equivariance to object changes, and utilizing additional sensor data for validation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If neural networks are used for object recognition in semi-automated driving, then the complexity of processing visual information is reduced and classification speed is improved, but the reliability of object recognition deteriorates due to lack of plausibility checks
Solution Approach 1:
The patent implements a feedback mechanism by computing relevance assessments from the neural network output and comparing them against expected relevance patterns. This feedback loop enables plausibility verification of classification results without reprocessing the original image data, thus maintaining high classification speed while improving reliability through automated consistency checks.
Solution Approach 2:
The patent introduces relevance assessments as an intermediary element between the neural network classification and the final object recognition decision. These relevance assessments serve as a mediator that provides additional verification information, allowing the system to check plausibility of classifications without directly reanalyzing the complex image data, thereby improving reliability while preserving processing efficiency.
2Reliability
If relevance assessment functions are applied to verify neural network outputs, then the reliability of classification is improved, but the computational complexity and processing time increase
Solution Approach 1:
The patent applies partial action by computing relevance assessments only for the specific classification outputs that need verification, rather than reprocessing the entire image recognition pipeline. This selective application of relevance assessment reduces computational overhead while maintaining reliability improvements for critical classification decisions.
Solution Approach 2:
The patent uses copying by generating relevance assessments from the existing neural network output data without requiring access to or reprocessing of the original image inputs. This copying approach allows plausibility verification to be performed on derived data structures, significantly reducing computational complexity compared to full image reanalysis.
3Measurement precision
If multiple feature parameters are determined for each image to enable plausibility checks, then the accuracy of relevance assessment is improved, but the quantity of data to be processed increases
Solution Approach 1:
The patent extracts only the essential feature parameters needed for plausibility verification from the full set of neural network outputs. By selecting and processing only the most relevant features for relevance assessment, the system achieves accurate verification without being burdened by the complete volume of raw classification data, thus improving measurement precision while controlling data quantity.
Data Source
AI summary
A method for a plausibility check of the output of an artificial neural network (ANN) utilized as a classifier. The method includes: a plurality of images for which the ANN has ascertained an association with one or multiple classes of a predefined classification, and the association that is ascertained in each case by the ANN, are provided; for each image at least one feature parameter is determined which characterizes the type, the degree of specificity, and/or the position of at least one feature contained in the image; for each combination of an image and an association, a spatially resolved relevance assessment of the image is ascertained by applying a relevance assessment function; a setpoint relevance assessment is ascertained for each combination, using the feature parameter; a quality criterion for the relevance assessment function is ascertained based on the agreement between the relevance assessments and the setpoint relevance assessments.