Object Detection Abnormality Screening With XAI Saliency Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models struggle to detect various types of abnormalities, including partial occlusion, out-of-distribution, and adversarial patches, as specialized algorithms are designed for specific targeted types, failing to recognize other forms of abnormalities.
Innovation Solution
An abnormality detection function that calculates a DiL score based on a saliency map and object detection results to quantify the effect of any type of abnormality on the decision-making process, using explainable AI techniques to measure inconsistency between intermediate and final stages of the object detection pipeline.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If specialized algorithms are designed for each individual type of abnormalities, then detection accuracy for targeted types is improved, but the ability to detect other types of abnormalities deteriorates
Solution Approach 1:
The patent applies universality by creating a single abnormality detection model that can detect multiple types of abnormalities (adversarial patches, natural anomalies, occlusions) simultaneously. The model uses a unified training approach with diverse training data containing various abnormality types, enabling one model to perform multiple detection functions rather than requiring separate specialized algorithms for each abnormality type.
2Measurement precision
If specialized algorithms are designed for specific targeted types of abnormalities, then detection precision for those types is improved, but the system becomes complex with multiple algorithms
Solution Approach 1:
The patent merges multiple specialized detection algorithms into a single unified abnormality detection model. By combining diverse training data containing various abnormality types and training a single model on this comprehensive dataset, the system consolidates multiple detection functions into one model, reducing system complexity while maintaining detection precision across different abnormality types.
3Productivity
If existing object detection models are used without abnormality detection function, then object detection speed is maintained, but the ability to detect abnormalities deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-training the object detection model with diverse training data that includes various types of abnormalities embedded within normal images. This preliminary training equips the model with abnormality detection capabilities before actual use, allowing it to simultaneously perform both object detection and abnormality detection without requiring separate processing steps, thereby maintaining detection speed while enhancing reliability.
Data Source
AI summary
A non-transitory computer-readable recording medium stores therein an abnormality detection program that causes a computer to execute a process including calculating a first value and a second value, based on a detection result of an object detected from an input scene to an object detection model and a saliency map for the input scene obtained by an XAI technique, the first value indicating a value of the saliency map for the entire input scene, the second value indicating a value of the saliency map in a region other than a region of the detected object in the input scene, and detecting an abnormality in the input scene based on the calculated first and second values.