Digital Image Anomaly Detection Using Conformal Class Subsets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital image processing methods for anomaly detection in autonomous driving suffer from performance degradation due to environmental changes, such as lighting and weather conditions, which affect the accuracy of anomaly detection.
Innovation Solution
A device and computer-implemented method that uses a trained model to process digital images by determining a subset of classes based on probability assignments, where the size of the subset indicates an anomaly score, and adjusts thresholds based on a weighted sum of scores to account for domain shifts in environmental conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a model is trained on digital images from multiple environmental conditions, then the anomaly detection accuracy is improved across different environments, but the training data requirements and model complexity increase
Solution Approach 1:
The patent segments the anomaly detection process into two distinct phases: (1) training phase using images from multiple environmental conditions to build a robust model, and (2) detection phase using conformal prediction with probability thresholds. This segmentation allows the system to learn from diverse environments during training while applying a simplified statistical approach during actual anomaly detection, thereby improving reliability without proportionally increasing operational complexity.
Solution Approach 2:
The patent performs preliminary actions by pre-training the model on a comprehensive data set containing images from various environmental conditions (different lighting, weather, times of day) before deployment. This preliminary training establishes a strong baseline model that can then use conformal prediction to adapt to new environments without requiring retraining, thus improving anomaly detection accuracy across different environments while keeping the operational model relatively simple.
2Measurement precision
If conformal prediction with probability thresholds is used to detect anomalies, then the detection accuracy is improved, but the computational complexity increases
Solution Approach 1:
The patent changes the parameter approach by using conformal prediction to dynamically determine probability thresholds based on the model's confidence scores. Instead of using fixed thresholds, the system calculates thresholds that guarantee a specified coverage level (e.g., 90% or 95% confidence) that the true class is included in the predicted set. This parameter adaptation improves detection precision while keeping computations manageable by leveraging the existing model outputs.
Solution Approach 2:
The patent substitutes traditional anomaly detection mechanisms (such as fixed thresholding or complex statistical tests) with conformal prediction, which provides a unified framework for uncertainty quantification. This substitution replaces ad-hoc detection methods with a statistically rigorous approach that guarantees coverage levels, improving precision without requiring complex additional computational infrastructure.
3Adaptability or versatility
If the model is trained on a diverse data set with different environmental conditions, then the adaptability to domain shifts is improved, but the data set size and processing time increase
Solution Approach 1:
The patent performs preliminary action by pre-training the model on a comprehensive data set that includes images from various environmental conditions (different lighting, weather, times of day, locations). This one-time extensive training establishes a robust model that can generalize to new environments. The conformal prediction component then enables the model to adapt to domain shifts without requiring additional training data or retraining, thus improving adaptability while limiting the time loss to the initial training phase only.
4Reliability
If the subset size is determined based on conformal prediction, then the anomaly detection reliability is improved, but the computational resources required increase
Solution Approach 1:
The patent changes the parameter determination approach by using conformal prediction to dynamically adjust the subset size based on the model's confidence scores and the desired coverage level. Instead of using a fixed subset size or complex statistical methods, the system calculates the minimum subset size required to guarantee the specified reliability level. This parameter adaptation improves anomaly detection reliability while optimizing computational energy consumption by avoiding unnecessary computations.
Data Source
AI summary
Device and computer implemented method for processing a digital image for anomaly detection. The method includes providing a model that is trained with a data set that includes digital images to output probabilities that are assigned to classes in a set of classes for classifying or semantically segmenting the digital image depending on the digital image; determining the probabilities for the digital image with the model; determining a size of a sub-set of the set of classes depending on a sum of the probabilities that are assigned to the classes in the sub set and depending on a first threshold; and detecting an anomaly depending on the size; wherein the sum of the probabilities is smaller than the first threshold, determining the size of the sub-set comprises determining the first threshold depending on a weighted sum of scores, a score is weighted in the weighted sum.