Hyperspectral Target Detection Using Momentum-Updated Binary Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classical hyperspectral target detection algorithms face challenges in separating background from targets due to spectral variation and mixed pixels, resulting in low detection rates and high false alarm rates, especially when prior spectral information is limited.
Innovation Solution
A hyperspectral target detection method using a binary-classification encoder network with momentum update, involving clustering, screening of pure pixels, constructing a background-target training sample set, and optimizing the network with a loss function to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If classical target detection algorithms (CEM, ACE, OSP, SRBBH, SLRMD) are used, then the detection process is simple and fast, but the detection rate is low and false alarm rate is high due to spectral variation and mixed pixel problems
Solution Approach 1:
The algorithm segments pixels into pure pixels and mixed pixels through clustering analysis. Pure pixels are used for training the deep learning model, while mixed pixels are handled separately through spectral unmixing techniques. This segmentation allows the model to focus on clean training data while still addressing the complexity of mixed pixels in the detection phase.
Solution Approach 2:
The patent transitions from traditional spectral space analysis to a deep feature space by constructing a binary-classification encoder network. The network extracts deep features from hyperspectral images and projects them into a new dimensional space where target and background are better separated, effectively adding a new dimension of analysis beyond conventional spectral methods.
2Reliability
If deep learning models are trained with sufficient samples, then the model effect and generalization ability are improved, but in practice only 1-5 pieces of prior spectral information are available making training conditions unmet
Solution Approach 1:
The algorithm performs preliminary clustering to identify and extract pure pixels from the hyperspectral image before training the deep learning model. This preliminary action creates a sufficient training dataset from the limited prior spectral information by generating multiple training samples through spectral unmixing and pixel combination techniques, enabling effective model training without requiring large amounts of pre-collected training data.
Solution Approach 2:
The system uses the hyperspectral image itself to generate training data through automated clustering and pure pixel extraction. The algorithm self-generates training samples by identifying pure pixels and creating synthetic mixed pixels through spectral unmixing, eliminating the need for external training datasets and enabling the model to learn from the image's own content.
3Quantity of substance
If prior spectral information is limited (1-5 pieces), then data collection is efficient, but the trained model effect is insufficient and detection accuracy deteriorates
Solution Approach 1:
The patent changes the parameters of training samples by generating multiple variations from limited prior spectral information. Through spectral unmixing, pixel combination, and augmentation techniques, the system transforms a small number of prior spectral samples into a large diverse training dataset with varying spectral compositions, enabling effective model training despite limited initial data.
Data Source
AI summary
A hyperspectral target detection method of a binary-classification encoder network based on a momentum update is provided, and includes following steps: converting an acquired 3-D hyperspectral image into a hyperspectral image in a 2-D matrix form, performing a clustering to obtain a clustering result, and initializing a centroid; based on the clustering result, using Euclidean distance to find pixels adjacent to each centroid as pure background pixels and target pixels, and screening pure pixels; constructing a background-target training sample set based on the pure pixels, constructing a binary-classification encoder network based on a momentum update through the background-target training sample set, calculating a loss function, and optimizing to obtain a trained binary-classification encoder network; inputting the hyperspectral image in the 2-D matrix form into the trained binary-classification encoder network, and outputting a final detection map.