SAR Target Recognition with Limited Training Samples in Open Set Conditions

Xiangyu Zhou, Yifan Zhang, Di Liu, Qianru WeiSensors 2023 · DOI

TL;DR

This paper tackles the challenging problem of recognizing Synthetic Aperture Radar (SAR) targets when the model encounters categories it wasn't trained on (open set recognition) and has very few training examples. They propose a novel Graph Convolutional Network (GNN)-based approach that builds a "relationship space" between target features to identify both known and unknown targets, achieving competitive accuracy and offering interpretability for unseen targets.

What Problem Does This Paper Address?

Collecting and labeling SAR imagery is notoriously difficult, leading to a common scenario in real-world applications: a limited number of training samples and categories. This is compounded by the "open set recognition" (OSR) problem, where a deployed system might encounter targets from categories it has never seen before. Traditional OSR methods, often designed for optical images and relying on abundant data or side information (like attributes), are not well-suited for SAR due to these constraints. The paper aims to bridge this gap by developing an OSR method specifically for SAR that is effective with limited training data and does not require external information about the unseen categories.

Methods

The core of the proposed method is a task-oriented OSR approach that leverages Graph Convolutional Networks (GNNs). Here's a breakdown:

  • Feature Extraction: A convolutional neural network (CNN) acts as a feature extractor, processing SAR images into low-dimensional feature vectors. Attention modules are incorporated to enhance important feature information.
  • Relationship Space Construction: The extracted features are then fed into a "relation space" built using a GNN. This space essentially models the relationships and similarities between different target features.
  • Graph Construction: A graph is constructed where nodes represent the features of samples in a given task. Edges represent the similarity between these features, calculated using a Multi-Layer Perceptron (MLP) that takes the Euclidean distance between feature vectors as input.
  • Iterative Refinement: The GNN employs multiple "adjacency modules" and "update modules" that iteratively refine the feature representations by considering their relationships within the graph. This process helps to learn a more robust representation of inter-category similarities and differences.
  • Classification: Finally, a fully-connected (FC) layer uses these refined features to output a probability distribution. For a query sample, if the highest probability is assigned to a "known" category, it's classified as such. If the model assigns a high probability to an "unseen" category (or a low probability to all known categories), it's flagged as unknown. The method also aims to interpret why a target is classified as unseen by analyzing its similarity to known categories.

Datasets

The primary dataset used for evaluation is the MSTAR dataset. This dataset is a standard benchmark for SAR target recognition and includes optical and SAR images of various ground targets under different conditions. The paper specifically mentions using the standard operating condition (SOC) with two depression angles (15° and 17°), encompassing 10 types of ground targets. For their experiments, they strategically divide categories into "seen" and "unseen" sets, with a portion of the seen categories also being used to simulate unseen categories during training to mimic real-world OSR scenarios.

Key Insights and Ideas

  • GNNs for SAR OSR: The paper demonstrates that GNNs are a powerful tool for SAR OSR, particularly in modeling the complex relationships between target features without relying on explicit side information.
  • Task-Oriented Learning: By training the model on a series of "tasks" (each with a support set and a query set), the model learns to adapt and construct the relationship space effectively even with limited samples per task. This is a key aspect of its few-shot learning capability.
  • Interpretability of Unseen Targets: Beyond just classifying a target as "unseen," the method provides insights into why by showing its similarity to known categories. This is crucial for understanding model behavior and building trust.
  • Limited Data is Realistic: The paper strongly emphasizes that the difficulty of SAR data collection makes limited training samples a realistic constraint, and their method is designed to address this directly.

Key Results

  • High Accuracy for Seen Targets: The proposed method consistently achieves high recognition accuracy for seen categories, often exceeding 95% across different classification problems (e.g., 99.06% for 3-class, 96.00% for 5-class).
  • Competitive Accuracy for Unseen Targets: While challenging, the method shows promising accuracy for unseen targets: 67.00% for a 3-type problem, 52.76% for a 5-type problem, and 34.12% for a 7-type problem (with K=20 samples).
  • Superiority over Baselines: Compared to methods like Prototype Network, Fea-DA, and GAN_OSR, the proposed GNN-based approach demonstrates better or more stable performance, especially under limited training data and open set conditions.
  • Effective Clustering: Visualization (t-SNE) shows that the proposed model effectively clusters samples of the same category together and separates different categories, a crucial aspect for OSR.

Limitations and Future Work

The authors acknowledge several limitations:

  • Performance Degradation with More Categories: The recognition accuracy tends to decrease as the number of categories increases.
  • Difficulty with Similar Unseen Categories: It is challenging to recognize an unseen category that is very similar to a seen category.
  • Single Unseen Category Identification: The current method is designed to identify a single unseen category at a time. It cannot handle scenarios where multiple distinct unseen categories are present simultaneously.

Future work will focus on addressing these limitations, particularly by developing methods to handle a larger number of categories, improve discrimination of similar unseen categories, and extend the framework to identify multiple unseen categories.

Summary updated

Abstract

It is difficult to collect training samples for all types of synthetic aperture radar (SAR) targets. A realistic problem comes when unseen categories exist that are not included in training and benchmark data at the time of recognition, which is defined as open set recognition (OSR). Without the aid of side-information, generalized OSR methods used on ordinary optical images are usually not suitable for SAR images. In addition, OSR methods that require a large number of samples to participate in training are also not suitable for SAR images with the realistic situation of collection difficulty. In this regard, a task-oriented OSR method for SAR is proposed by distribution construction and relation measures to recognize targets of seen and unseen categories with limited training samples, and without any other simulation information. The method can judge category similarity to explain the unseen category. Distribution construction is realized by the graph convolutional network. The experimental results on the MSTAR dataset show that this method has a good recognition effect for the targets of both seen and unseen categories and excellent interpretation ability for unseen targets. Specifically, while recognition accuracy for seen targets remains above 95%, the recognition accuracy for unseen targets reaches 67% for the three-type classification problem, and 53% for the five-type classification problem.
BibTeX
@inproceedings{zhou2023sar,
  author = {Xiangyu Zhou and Yifan Zhang and Di Liu and Qianru Wei},
  title = {{SAR Target Recognition with Limited Training Samples in Open Set Conditions}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23031668},
}
Figure