Global-Local Gaussian Modeling for Unsupervised Image Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods in images face challenges due to the lack of a clear definition of normality and anomalies, especially in unsupervised scenarios, leading to inefficiencies and data constraints that limit the effectiveness of current models, particularly in industrial applications.
Innovation Solution
A method utilizing a global-local Gaussian mixture model that combines the flexibility of local modeling with the robustness of a global Gaussian mixture, leveraging pre-trained neural networks to extract features and apply a K-MLE algorithm for online learning, and incorporating local weight maps to enhance anomaly detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are used for anomaly detection, then implementation is straightforward, but they fail to capture complex non-linear relationships and require extensive hyperparameter tuning
Solution Approach 1:
The model is segmented into two distinct components: a global model that captures overall data patterns and a local model that focuses on specific anomaly patterns. This segmentation allows each component to specialize, improving detection accuracy while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent combines multiple neural network architectures (CNN for spatial features, RNN for temporal features) into a composite model structure. This composite approach leverages the strengths of different architectures to capture complex non-linear relationships while providing a structured framework that manages overall model complexity.
2Measurement precision
If extensive hyperparameter tuning is performed to improve detection accuracy, then anomaly detection precision improves, but computational cost and time consumption increase
Solution Approach 1:
The model performs preliminary feature extraction and representation learning during the training phase, transforming raw data into meaningful feature embeddings. This preliminary action reduces the complexity of subsequent anomaly detection, lowering computational costs during deployment while maintaining high detection precision.
Solution Approach 2:
The patent uses autoencoders to learn and store normal data patterns during training, creating a compressed representation of normal behavior. During detection, the model compares actual data against this learned copy of normal patterns, reducing computational requirements compared to analyzing raw data from scratch.
3Measurement precision
If the model focuses on specific local patterns, then detection precision for certain anomalies improves, but the model becomes biased towards specific patterns and misses broader anomalies
Solution Approach 1:
The patent merges a global model that captures overall data distributions with local models that specialize in specific anomaly patterns. This combination allows the system to maintain adaptability through the global model while achieving high precision for specific patterns through local models, resolving the trade-off between specialization and versatility.
Solution Approach 2:
The global model serves as a universal component that handles diverse and unknown anomaly types, while local models provide specialized detection for known patterns. This multi-functional architecture ensures the system remains adaptable to new anomaly types while maintaining high precision for recognized patterns.
Data Source
Figure 1
AI summary
We propose a global-local Gaussian model for features of images produced by the intermediate layers of a network of neurons of the neural network to detect anomalies without external supervision. A global Gaussian mixture model of features (or descriptors) is first of all learned using all available features of normal data. This global Gaussian mixture model is then localized by learning local weights for each Gaussian. Finally, these weights are used instead of those of the mixture to detect anomalies. This method allows precise modelling of complex data, even with limited data. Applied to the WideResnet 50-2 features, our approach surpasses the prior art on the MVTec dataset, in particular on the category of objects. It is robust to frequent disturbances on production lines, such as imperfect alignment, and is comparable to the prior art in terms of memory and computing time.