Anomaly Detection via Segmented Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing models for anomaly detection in physical products are often inaccurate due to the variety of anomaly types, characteristics, and environmental changes, leading to false negative or false positive detections, as they are not trained to recognize specific anomalies or their characteristics.
Innovation Solution
An object analysis system utilizing a combination of a feature extraction model, classification model, and anomaly localization model trained on reference images of non-anomalous objects, allowing for robust and accurate detection, classification, and localization of anomalies without prior training on specific anomaly types or characteristics, while conserving computing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing image processing models are used for anomaly detection, then the detection process can be performed, but the accuracy is low due to false positive and false negative detections
Solution Approach 1:
The patent segments the anomaly detection task into three distinct model components: a feature extraction model that identifies object features, a classification model that determines anomaly status, and an anomaly localization model that pinpoints anomaly locations. This segmentation allows each model to specialize in a specific aspect of detection, improving overall accuracy and reducing false positives and negatives compared to monolithic existing models.
Solution Approach 2:
The patent performs preliminary action by training all three models exclusively on reference images of non-anomalous objects before deployment. This preliminary training establishes a baseline understanding of normal object characteristics, enabling the models to effectively identify deviations as anomalies without requiring training on diverse anomaly types, thereby improving detection reliability.
2Adaptability or versatility
If models are trained on diverse anomaly types to improve detection coverage, then more anomaly types can be detected, but the training data requirements and computing resources increase
Solution Approach 1:
The patent inverts the conventional training approach by training models on non-anomalous objects rather than on anomaly examples. This inversion allows the system to achieve versatile anomaly detection coverage without requiring training data for each specific anomaly type. The models learn what normal objects should look like and automatically identify any deviations as anomalies, regardless of the specific anomaly type.
Solution Approach 2:
The patent creates universal models that can detect multiple types of anomalies through a single training process on non-anomalous objects. The feature extraction model, classification model, and localization model collectively provide multi-functional capability to detect various anomaly types (defects, contaminants, structural issues) without requiring separate trained models for each anomaly category, thus reducing training data requirements while maintaining adaptability.
Data Source
AI summary
An object analysis system is disclosed herein. The object analysis system may receive an input image that depicts an object. The object analysis system may determine, using a feature extraction model and from the input image, a first feature output that is associated with one or more features of the object. The feature extraction model may be trained based on reference images that depict reference objects that are a type of the object. The object analysis system may determine, using a classification model, that an anomaly status of the object is indicative of the object including an anomaly. The classification model may be trained based on the reference images. The object analysis system may determine, using an anomaly localization model, a location of the anomaly in the input image based on a second feature output of the convolutional neural network encoder. The anomaly localization model may be trained based on the reference images. The object analysis system may perform an action associated with the location of the anomaly.