Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Assignees
Member
NoblisNoblis is a nonprofit research and technical organization supporting federal missions in defense, health, environment, and security. Emphasizing applied sciences, engineering, digital transformation, artificial intelligence, cloud, and cybersecurity, Noblis provides objective solutions for government agencies confronting complex operational and scientific challenges.
Noblis is a nonprofit research and technical organization supporting federal missions in defense, health, environment, and security. Emphasizing applied sciences, engineering, digital transformation, artificial intelligence, cloud, and cybersecurity, Noblis provides objective solutions for government agencies confronting complex operational and scientific challenges.
Publication Number
US-11645875-B2
Publication Date
2023-05-09
Expiration Date
Abstract
Techniques for detecting anomalies in multispectral image data, and more specifically for detecting presentation attacks by using multispectral image data in biometric security applications, are provided. In some embodiments, a system may receive multispectral image data and generate an estimation of a first image of a plurality of images of the multispectral image data, wherein the estimation is based on other images of the multispectral image data, but not the first image itself. The estimation may then be compared to the first image to generate an indication as to whether the multispectral image data represents a presentation attack. In some embodiments, a system may receive multispectral training image data and may extract features from the data to generate and store a network architecture for predicting relationships of multispectral images of subjects.
Core Innovation
The invention generates and stores network architectures for biometric security using multispectral training image data. The training data includes, for each subject of a plurality of subjects, multispectral image data comprising a plurality of images of the subject, where each image is at a different wavelength range. The method aligns the corresponding plurality of images for each subject into a respective multispectral cube representing the multispectral image data for the respective subject.
From the aligned multispectral cube data, the invention extracts features from the multispectral training image data. Based on the extracted features, it generates and stores a network architecture for predicting relationships of multispectral images of a subject. The network architecture comprises an auto-encoder network architecture comprising a plurality of convolutional neural networks.
The generated auto-encoder-based convolutional neural network architecture is configured to generate an estimation of a first image of a subject at a first wavelength range. The estimation is generated based on a plurality of images of the subject at respective plurality of wavelength ranges different from the first wavelength range, thereby learning cross-wavelength relationships and producing an estimated image at the first wavelength range.
Claims Coverage
The independent claims in the provided set are clm-00001, clm-00010, and clm-00011. Across these independent claims, the coverage centers on aligning multispectral images into multispectral cubes, extracting features, and generating and storing an auto-encoder network architecture comprising multiple convolutional neural networks that estimates an image at one wavelength range from images at other wavelength ranges.
Multispectral cube alignment from multi-subject multi-wavelength image sets
Receiving multispectral training image data comprising, for each subject of a plurality of subjects, multispectral image data comprising a plurality of images of the subject, each of the plurality of images for each subject being an image at a different wavelength range; and for each of the plurality of subjects, aligning the corresponding plurality of images into a respective multispectral cube representing the multispectral image data for the respective subject.
Feature extraction and generating an auto-encoder convolutional network architecture
Extracting features from the multispectral training image data; and generating and storing, based on the extracted features, a network architecture for predicting relationships of multispectral images of a subject, wherein the network architecture comprises an auto-encoder network architecture comprising a plurality of convolutional neural networks.
Cross-wavelength image estimation as predicted relationships
Configuring the network architecture to generate an estimation of a first image of a subject at a first wavelength range based on a plurality of images of the subject at a respective plurality of wavelength ranges different from the first wavelength range.
Siamese ensemble likelihood for cross-wavelength consistency (dependent)
Using a Siamese ensemble of convolutional neural networks to estimate the likelihood that a subject’s first image at one wavelength range matches a subject’s second image at another wavelength range.
Regional patch-based feature extraction (dependent)
Extracting features from the multispectral training image data by dividing the images into multiple regional patches.
Wavelet filter bank feature extraction (dependent)
Extracting features from the multispectral training image data using a wavelet filter bank with multiple filters.
Optimized subset of wavelet filters for predetermined wavelength pairs (dependent)
Using only a selected subset of filters chosen to optimize performance for a predetermined pair of wavelength ranges.
Tensor-dictionary multilinear mapping feature extraction (dependent)
Extracting features from the multispectral training image data by using tensor dictionaries that provide multilinear mappings across a set of vector spaces.
Across the independent claims, the essential inventive structure is aligning each subject’s multispectral images into a multispectral cube, extracting features, and generating and storing an auto-encoder network architecture comprising multiple convolutional neural networks that estimates an image at a first wavelength range from images at other wavelength ranges. The dependent claim set further specifies refinements to feature extraction and adds a Siamese ensemble likelihood for cross-wavelength consistency.
Stated Advantages
Generates an estimation of a first image of a subject at a first wavelength range based on images at different wavelength ranges.
Predicts relationships of multispectral images of a subject using an auto-encoder network architecture comprising convolutional neural networks.
Documented Applications
Multispectral presentation-attack detection (PAD) via cross-spectral analysis, including using hallucinated expected images and comparing with genuine images to determine anomalies.
Biometric security for modalities in a dataset context including face, iris, and fingerprint.
Interested in licensing this patent?