Hyperspectral Anomaly Detection Using Local and Global Autoencoders
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional hyperspectral anomaly detection methods struggle to distinguish normal and abnormal data due to noise and generalization issues, especially when abnormal data is similar to normal data, and fail to effectively utilize local noise in hyperspectral imaging.
Innovation Solution
Anomaly detection device using a neural network that extracts local features through partial autoencoders and global features through an aggregate autoencoder, with fine-tuning of a target neural network based on trained partial autoencoders to enhance anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional anomaly detection methods are used in hyperspectral imaging, then the detection process is simple, but the ability to distinguish normal and abnormal data deteriorates due to noise and generalization issues
Solution Approach 1:
The patent divides the hyperspectral image into multiple local regions and processes each region through separate partial autoencoders. This segmentation allows the system to capture local features and noise patterns in different regions, improving anomaly detection accuracy by comparing local variations against global patterns from the aggregate autoencoder.
Solution Approach 2:
The patent implements a nested architecture where multiple partial autoencoders (handling individual bands) are embedded within an aggregate autoencoder (handling multiple bands collectively). The partial autoencoders extract local features from individual bands, which are then fed into the aggregate autoencoder to capture global patterns, creating a hierarchical feature extraction system that resolves the contradiction between detection accuracy and system complexity.
2Measurement precision
If partial autoencoders are used to extract local features, then the ability to detect local anomalies improves, but the computational complexity increases
Solution Approach 1:
The patent applies partial action by using multiple partial autoencoders that each process only a subset of bands locally, rather than processing all bands globally. This partial processing reduces the computational burden on each individual autoencoder while collectively capturing comprehensive local features across all bands through the aggregate autoencoder.
Solution Approach 2:
The patent merges the outputs of multiple partial autoencoders into an aggregate autoencoder that processes multiple bands collectively. This combining approach allows the system to leverage local feature extraction from partial autoencoders while maintaining computational efficiency through shared processing in the aggregate autoencoder, balancing local accuracy with overall computational power consumption.
3Measurement precision
If fine-tuning is performed on target neural network, then the anomaly detection accuracy improves, but the training time increases
Solution Approach 1:
The patent performs preliminary training of partial autoencoders on individual bands before fine-tuning the aggregate autoencoder with all bands. This preliminary action allows the system to establish initial feature extraction capabilities in parallel for each band, reducing the overall training time while maintaining the ability to perform subsequent fine-tuning for improved anomaly detection accuracy.
Solution Approach 2:
The patent implements a dynamic training approach where the system transitions from static pre-training of partial autoencoders to dynamic fine-tuning of the aggregate autoencoder. This dynamic process allows the system to adaptively adjust parameters based on preliminary learning results, improving anomaly detection accuracy while managing training time through staged optimization rather than simultaneous full training.
Data Source
AI summary
An anomaly detection device according to one embodiment may comprise: a receiver for receiving a hyperspectral image; and a processor for extracting, on the basis of a plurality of target partial autoencoders corresponding to a plurality of bands included in the hyperspectral image, a plurality of local features corresponding to the plurality of bands, extracting a global feature on the basis of the plurality of local features through an aggregate autoencoder, and detecting anomalies on the basis of the global feature.