Apparatus and process for electromagnetic imaging
View Patent ↗A computer-implemented process for electromagnetic imaging, the process including the steps of: accessing scattering data representing, for each of a plurality of different electromagnetic wave energies, at least a corresponding two-dimensional array of measurements of electromagnetic wave scattering by internal features of an object, each said measurement representing scattering of electromagnetic waves emitted by a corresponding antenna of an array of antennas disposed about the object, scattered within the object, and measured by a corresponding antenna of the array of antennas; processing the scattering data with a plurality of trained neural networks to generate corresponding compressed scattering data, wherein the trained neural networks process respective subsets of the scattering data corresponding to respective different separations between the emitting and measuring antennas; processing the compressed scattering data with a trained neural network to generate corresponding compressed dielectric distribution data; and processing the compressed dielectric distribution data with a trained decompressor to generated dielectric distribution data representing a spatial distribution of a dielectric property within the object.
1 . A computer-implemented process for electromagnetic imaging, the process comprising:
accessing scattering data representing, for each of a plurality of different electromagnetic wave energies, at least a corresponding two-dimensional array of measurements of electromagnetic wave scattering by internal features of an object, each measurement representing scattering of electromagnetic waves emitted by a corresponding antenna of an array of antennas disposed about the object, scattered within the object, and measured by a corresponding antenna of the array of antennas;
processing the scattering data with a plurality of trained neural networks to generate corresponding compressed scattering data, wherein the trained neural networks process respective subsets of the scattering data corresponding to respective different separations between the emitting and measuring antennas;
processing the compressed scattering data with a trained neural network to generate corresponding compressed dielectric distribution data; and
processing the compressed dielectric distribution data with a trained decompressor to generated dielectric distribution data representing a spatial distribution of a dielectric property within the object;
wherein the trained neural networks are trained by unsupervised learning using training scattering data, the trained decompressor is trained by unsupervised learning using training dielectric distribution data representing spatial distributions of the dielectric property of objects, and the trained neural network that processes the compressed scattering data is trained by supervised learning using labelled training data.
2 . A computer-implemented process of claim 1 , wherein the processing of the compressed scattering data includes applying a calibration to the compressed scattering data to generate calibrated compressed scattering data, and the trained neural network generates the compressed dielectric distribution data from the calibrated compressed scattering data.
3 . The computer-implemented process of claim 1 , further comprising:
training the trained neural networks by unsupervised learning using the training scattering data;
training the trained decompressor by unsupervised learning using the training dielectric distribution data; and
training the neural network that processes the compressed scattering data by supervised learning using the labelled training data.
4 . The computer-implemented process of claim 1 , wherein the trained neural networks are one-dimensional convolution compressors of a scattering data auto-encoder that includes a ‘long short-term memory’ (“LSTM”) decompressor.
5 . The computer-implemented process of claim 1 , wherein the trained decompressor is a component of a variational auto-encoder.
6 . The computer-implemented process of claim 1 , wherein the trained neural network that processes the compressed scattering data is a single-layer neural network with no activation function.
7 . A computer-readable storage medium having stored thereon executable instructions that, when executed by at least one processor, cause the at least one processor to perform the process of claim 1 .
8 . An apparatus for electromagnetic imaging, including components configured to perform the process of claim 1 .
9 . An apparatus for electromagnetic imaging, comprising:
a scattering data compressor configured to access scattering data representing, for each of a plurality of different electromagnetic wave energies, at least a corresponding two-dimensional array of measurements of electromagnetic wave scattering by internal features of an object, each said measurement representing scattering of electromagnetic waves emitted by a corresponding antenna of an array of antennas disposed about the object, scattered within the object, and measured by a corresponding antenna of the array of antennas, and to process the scattering data using a plurality of trained neural networks to generate corresponding compressed scattering data, wherein the trained neural networks process respective subsets of the scattering data corresponding to respective different separations between the emitting and measuring antennas;
a mapping component configured to process the compressed scattering data with a trained neural network to generate corresponding compressed dielectric distribution data; and
a trained decompressor configured to process the compressed dielectric distribution data to generate dielectric distribution data representing a spatial distribution of a dielectric property within the object;
wherein the trained neural networks are trained neural networks of a scattering data auto-encoder configured to train the neural networks by unsupervised learning, the decompressor is a decompressor of a dielectric distribution auto-encoder configured to train the decompressor by unsupervised learning, and the trained neural network that processes the compressed scattering data is trained by supervised learning.
10 . An apparatus as claimed in claim 9 , wherein the mapping component is configured to apply a calibration to the compressed scattering data to generate calibrated compressed scattering data, and the trained neural is configured to generate the compressed dielectric distribution data from the calibrated compressed scattering data.
11 . The apparatus of claim 9 , wherein the trained neural networks of the scattering data auto-encoder are one-dimensional convolution compressors, and the scattering data auto-encoder includes a ‘long short-term memory’ (“LSTM”) decompressor.
12 . The apparatus of claim 9 , wherein the dielectric distribution auto-encoder is a variational auto-encoder.
13 . The apparatus of claim 9 , wherein the trained neural network of the mapping component is a single-layer neural network with no activation function.