Autonomous Vehicle Vision Model Degradation Detection by Pixel Entropy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting performance degradation in deep learning models for autonomous vehicles are time-consuming and inefficient, often failing to detect degradation under undefined conditions.
Innovation Solution
An apparatus and method that determine entropy of each pixel in an output image of a deep learning model, calculate an average entropy value, and detect performance degradation based on this value, allowing for quick and accurate detection without setting specific conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional methods set specific conditions one by one to detect performance degradation, then detection can be performed under defined scenarios, but it takes a lot of time and fails to detect degradation under undefined conditions
Solution Approach 1:
The patent changes the detection parameter from specific condition-based error codes to entropy values calculated from output images. By calculating entropy for each pixel in the output image and comparing it against threshold values, the system can detect performance degradation without setting specific conditions, thereby reducing detection time while maintaining reliability
Solution Approach 2:
The system uses the deep learning model's own output images to detect its performance degradation. By calculating entropy from the model's output without requiring external test data or manual condition setting, the model essentially monitors itself, eliminating the need for separate verification datasets and reducing detection time
2Reliability
If specific conditions are set for detecting performance degradation, then detection can be performed systematically, but the device complexity increases due to multiple condition settings
Solution Approach 1:
The patent extracts the essential characteristic of performance degradation (entropy change in output images) from complex condition settings. By focusing only on entropy calculation from output images rather than setting multiple specific conditions, the system reduces complexity while maintaining comprehensive detection coverage
Solution Approach 2:
The entropy-based detection method serves as a universal indicator for performance degradation across different scenarios and conditions. A single entropy calculation mechanism replaces multiple condition-specific detection rules, making the system simpler while maintaining broad applicability
Data Source
AI summary
An apparatus and method for detecting performance degradation of an autonomous vehicle include a storage configured to store a deep learning model used for image recognition in the autonomous vehicle and a controller configured to determine entropy of each pixel among a plurality of pixels in an output image of the deep learning model, determine an average value of the entropy of the plurality of pixels, and detect performance degradation of the deep learning model based on the average value of the entropy of the plurality of pixels.