Deep learning-based multiband signal fusion for 3-D SAR super-resolution
AssigneeBoard of Regents, The University of Texas System
InventorsJosiah W. Smith, Murat Torlak
Life of this patent
Estimating the term…
Abstract
Multiband radar fusion is provided. The method comprises collecting first radar data from a first radar having a first frequency and collecting second radar data from a second radar having a second frequency different from the first frequency. A hybrid, dual-domain complex value convolutional neural network (CV-CNN), fuses the first and second radar data. The CV-CNN alternates between operating in the wavenumber-domain and the spatial-domain of the first and second radar data. A synthetic aperture radar image is reconstructed from the fused first and second radar data according to Fourier-based algorithm.
Claims18
A computing-implemented method of multiband radar fusion, the method comprising:
- collecting first radar data from a first radar having a first frequency;
- collecting second radar data from a second radar having a second frequency different from the first frequency;
- fusing, by a neural network, the first and second radar data, wherein the neural network alternates between operating in the wavenumber-domain and the spatial-domain of the first and second radar data, wherein the neural network comprises:
- a first complex-valued convolution layer;
- four residual blocks following the first complex-valued convolution layer;
- domain transformation blocks interleaved between the four residual blocks;
- a pass-through connection in the wavenumber-domain from the first complex-valued convolution layer to after the fourth residual blocks;
- a fifth residual block following the pass-through connection;
- and
- a second complex-valued convolution layer;
- and
- reconstructing a synthetic aperture radar image from the fused first and second radar data according to Fourier-based algorithm.
A neural network for fusing radar data, comprising:
- a first complex-valued convolution layer;
- four residual blocks following the first complex-valued convolution layer;
- domain transformation blocks interleaved between the four residual blocks;
- a pass-through connection in the wavenumber-domain from the first complex-valued convolution layer to after the fourth residual blocks;
- a fifth residual block following the pass-through connection;
- and
- a second complex-valued convolution layer.
A multiband imaging system, comprising:
- two-axis mechanical scanner;
- a first radar having a first frequency mounted to the two-axis mechanical scanner;
- a second radar having a second frequency different from the first frequency mounted to the two-axis mechanical scanner;
- and
- a computer system connected to the first radar, second radar, and two-axis controller, wherein the computer system is configured to:
- collect first radar data from the first radar;
- collect second radar data from a second radar;
- fuse, by a neural network, the first and second radar data, wherein the neural network alternates between operating in the wavenumber-domain and the spatial-domain of the first and second radar data, wherein the neural network comprises:
- a first complex-valued convolution layer;
- four residual blocks following the first complex-valued convolution layer;
- domain transformation blocks interleaved between the four residual blocks;
- a pass-through connection in the wavenumber-domain from the first complex-valued convolution layer to after the fourth residual blocks;
- a fifth residual block following the pass-through connection;
- and
- a second complex-valued convolution layer;
- and
- reconstruct a synthetic aperture radar image from the fused first and second radar data according to Fourier-based algorithm.
Dependencies are inferred from the claim text.