SAR Classifier Neural Network for Onboard Target Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Converting synthetic aperture radar (SAR) range profiles into images for automatic target recognition (ATR) is computationally expensive, making it unsuitable for onboard processing in lightweight computing devices, and existing machine learning models require large amounts of training data, which is unavailable for certain types of objects, leading to inaccurate classification.
Innovation Solution
A computing system with a SAR classifier neural network that includes a SAR encoder, an image encoder, and a classifier, trained using both SAR range profiles and two-dimensional images, allowing for shared latent representations and transfer learning to classify objects with limited SAR data, enabling accurate classification without converting SAR range profiles into images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If SAR range profiles are converted into SAR images for ATR, then object recognition capability is improved, but computational cost increases significantly
Solution Approach 1:
The patent extracts the essential feature representation capability from the full SAR image generation process. Instead of converting entire SAR range profiles to images, the system extracts and uses only the essential spectral features directly from the range profile domain, eliminating the computationally expensive imaging step while retaining recognition capability.
Solution Approach 2:
The patent replaces the traditional mechanical/optical image formation process with a direct spectral feature extraction approach. Instead of using the SAR system to physically form images through radar signal processing, the system substitutes this with a computational approach that directly processes spectral features from range profiles, significantly reducing computational requirements.
2Reliability
If existing machine learning models are used for SAR ATR, then classification capability is achieved, but large amounts of training data are required
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing spectral feature representations from SAR range profiles during the training phase. These pre-extracted spectral features serve as compressed, informative representations that capture essential object characteristics, allowing the machine learning model to learn effective classification patterns without requiring vast amounts of raw SAR data.
Solution Approach 2:
The patent changes the parameter representation from raw SAR range profile data to transformed spectral feature representations. By converting the input parameters from the original SAR domain to a spectral feature domain, the system reduces the effective dimensionality and redundancy in the data, enabling better classification performance with smaller training datasets.
3Power
If SAR range profiles are processed directly without image conversion, then computational cost is reduced, but classification accuracy may deteriorate
Solution Approach 1:
The patent applies local quality by focusing the feature extraction on specific spectral regions and characteristics that are most informative for object recognition. Instead of processing the entire SAR range profile uniformly, the system identifies and extracts local spectral features that capture essential object signatures, maintaining classification accuracy while reducing overall computational burden.
Data Source
AI summary
A computing system including a processor configured to train a synthetic aperture radar (SAR) classifier neural network. The SAR classifier neural network is trained at least in part by, at a SAR encoder, receiving training SAR range profiles that are tagged with respective first training labels, and, at an image encoder, receiving training two-dimensional images that are tagged with respective second training labels. Training the SAR classifier neural network further includes, at a shared encoder, computing shared latent representations based on the SAR encoder outputs and the image encoder outputs, and, at a classifier, computing respective classification labels based on the shared latent representations. Training the SAR classifier neural network further includes computing a value of a loss function based on the plurality of first training labels, the plurality of second training labels, and the plurality of classification labels and performing backpropagation based on the value of the loss function.