Abstract
Synthetic Aperture Radar (SAR) target recognition is a critical component of SAR image interpretation. However, the high cost and specialized expertise required for labeling SAR images often limit the effectiveness of supervised learning methods. This paper presents an end-to-end semi-supervised recognition method that leverages an attention mechanism and bias-variance decomposition to intelligently screen unlabeled data. Central to this approach is the Dataset Attention Module (DAM), which prioritizes informative unlabeled samples while filtering out noisy or hard-to-learn data. Experimental results on the MSTAR dataset demonstrate that our method significantly outperforms state-of-the-art semi-supervised algorithms, achieving over 93% accuracy even when 90% of the data is unlabeled.
Introduction
Synthetic Aperture Radar (SAR) provides high-resolution Earth observation data regardless of weather or lighting conditions. Despite its utility, SAR images are difficult for humans to interpret due to speckle noise and orientation sensitivity, making manual labeling a bottleneck for deep learning applications. Semi-supervised learning (SSL) offers a solution by utilizing small labeled datasets alongside larger unlabeled ones. However, the quality of unlabeled data is often inconsistent, and incorrect "pseudo-labels" can degrade model performance. This research addresses these challenges by introducing a framework that automatically evaluates and selects the most beneficial unlabeled samples for training.
Methodology
The proposed framework operates in two distinct stages to ensure robust learning from limited labeled data.
Stage 1: Baseline Training and Uncertainty Collection
Initially, a baseline model is trained using only the available labeled samples. This model utilizes a ResNeXt architecture enhanced with a Convolutional Block Attention Module (CBAM). The CBAM-CNN architecture integrates both channel and spatial attention to focus on relevant features within the SAR images, effectively ignoring background clutter.
Once the baseline model is stable, it is used to predict labels for the unlabeled dataset multiple times. These multiple predictions allow for the calculation of bias (accuracy) and variance (consistency) for each sample. Samples with low bias and low variance are identified as reliable candidates for pseudo-labeling.
Stage 2: The Dataset Attention Module (DAM)
The core innovation of this work is the Dataset Attention Module (DAM). Unlike traditional attention mechanisms that operate within a single image, the DAM applies attention to the dataset itself. The overall two-step process is illustrated in Figure 4.
The DAM groups unlabeled data into sub-datasets and assigns a Dataset Screening Factor (DSF) to each. Groups that contribute to a reduction in overall training loss are assigned higher DSFs, ensuring that the model focuses on the most informative samples in subsequent iterations. This iterative filtering process mitigates the risk of single-prediction errors and improves the overall quality of the training set.
Experimental Results
The method was evaluated using the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset, the standard benchmark for SAR target recognition. The dataset includes 10 categories of vehicles, such as tanks and artillery.
The proposed DA-CBAM method demonstrated exceptional performance across various unlabeled rates:
- High Efficiency: At a 90% unlabeled rate (only 275 labeled images for 10 classes), the model achieved 93.20% accuracy, significantly higher than traditional SSL methods like the Ladder-network (66.58%).
- Robustness: The model showed increased resilience to speckle noise compared to the baseline, indicating that the DAM-integrated filtering process helps the network learn more fundamental, stable features.
- Target Focus: Visualizations using Grad-CAM confirmed that the DA-CBAM model focuses more precisely on the actual vehicle targets rather than background noise.
Conclusion
By combining dataset-level attention with bias-variance decomposition, this research provides a powerful framework for SAR image recognition with minimal labeled data. The Dataset Attention Module effectively bridges the gap between supervised and semi-supervised learning by intelligently managing the quality of unlabeled samples. Future work will focus on optimizing hyperparameters and exploring more lightweight architectures to enable real-time deployment on edge devices like satellites and drones.
Summary updated