Neural Classifier Output Plausibility Using Relevance Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The reliability of object detection by neural classifier networks in automotive applications is compromised due to the lack of a robust method to validate the correctness of image evaluations, leading to potential misclassification of traffic-relevant objects, especially in complex scenarios or under adverse conditions.
Innovation Solution
A method is introduced to plausibilize the output of neural networks by using a relevance evaluation function to assess the contribution of image areas to classification decisions, with a secondary classifier trained to evaluate the agreement between reconstructions and actual assignments, and determining a figure of merit to select the most credible relevance evaluation function, which is then used to quantify the plausibility of the output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If neural networks are used for object classification in complex images, then classification capability is improved, but reliability of detection deteriorates due to inability to validate correctness
Solution Approach 1:
The patent introduces a relevance evaluation function as an intermediary between the neural network classifier and the final classification output. This function generates a relevance map that evaluates which image areas contribute to the classification decision, providing a validation mechanism that improves detection reliability without compromising classification capability
Solution Approach 2:
The system implements feedback by using the relevance evaluation results to plausibilize the neural network output. The relevance map provides information about the confidence and validity of classifications, creating a feedback loop that enhances reliability while maintaining the network's adaptability
2Measurement precision
If relevance evaluation is performed for all image areas, then measurement precision of classification is improved, but device complexity increases
Solution Approach 1:
The patent extracts only the relevant information from the complex neural network processing by generating a relevance map that highlights contributing image areas. This extraction approach provides precise classification validation without requiring the full complexity of the neural network to be directly involved in the validation process
3Ease of operation
If visual sampling is used to check classification correctness, then ease of operation is improved, but measurement precision deteriorates due to limited coverage
Solution Approach 1:
The system performs self-validation through automated relevance evaluation. The relevance map automatically assesses which image areas contribute to classification decisions, eliminating the need for manual visual sampling while providing comprehensive coverage and precise validation of all classifications
Data Source
AI summary
A method for plausibilizing the output of an artificial neural network (ANN) used as classifier. The method includes the following steps: a plurality of images, for which the ANN has determined an assignment to one or more classes of a predetermined classification, as well as the assignment determined in each case by the ANN are provided; for each combination of one image and one assignment, a location-resolved relevance evaluation of the image is determined utilizing a relevance evaluation function, this relevance evaluation indicating which parts of the image have contributed, to what extent, to the assignment; a further classifier is trained to determine from one image and one relevance evaluation ascertained for the image, a reconstruction of the assignment to which this relevance evaluation relates; based on the agreement between the reconstructions and the actual assignments, a figure of merit is determined for the relevance evaluation function.