Neural Network Training with Outlier-Guided Hyperspectral Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training neural network models with hyperspectral data, which contain many similar spectra, leads to repetitive training and significant time consumption due to duplicate spectral information.
Innovation Solution
A method and device for training neural networks that select an initial sample group, determine a final sample group by adding distribution difference data, calculate a final loss function based on distribution differences, and apply this function to reduce training time while maintaining performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all hyperspectral data is used for training the neural network model, then the training data comprehensiveness is improved, but the training time increases significantly
Solution Approach 1:
The patent extracts distribution difference data from the population by identifying outliers using outlier scores. This extraction process separates the valuable diverse spectral information from the redundant similar spectra, creating a refined training dataset that maintains comprehensiveness while reducing size and training time
Solution Approach 2:
Instead of using the complete population or a simple random sample, the patent applies partial action by selectively including only the most valuable samples - those with high outlier scores that represent distribution differences. This partial selection achieves better than full population training efficiency by focusing computational resources on the most informative data points
2Loss of time
If a simple random sample is used for training, then the training time is reduced, but the training performance deteriorates due to missing diverse spectral information
Solution Approach 1:
The patent changes the sampling parameter from random selection to selection based on outlier scores. By transforming the selection criterion from uniform probability to score-based ranking, the method identifies and selects samples that best represent the population distribution, thereby maintaining training performance while reducing the number of training samples needed
Solution Approach 2:
The patent replaces the mechanical random sampling process with an intelligent selection process that uses outlier score calculation and distribution analysis. This substitution transforms blind random selection into targeted sample selection, ensuring that the reduced training set captures the essential spectral diversity of the population
Data Source
AI summary
Disclosed is an anomaly detection device including a receiver configured to receive a hyperspectral image and a processor configured to detect outlier of the hyperspectral image based on a pretrained neural network, wherein a neural network includes a partial encoder to which each of a plurality of windows generated from the hyperspectral image is input; an aggregate autoencoder to which an output of each of the partial encoders is concatenated and input; and a partial decoder to which at least one of a divided output that is divided from an output of the aggregate autoencoder and the output of the partial encoder is input, and the neural network is trained based on a loss function including at least one of an intermediate output error for an intermediate output of the aggregate autoencoder and a partial reconstruction error for an output of the partial decoder.