Abstract
Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) is a critical field in remote sensing, yet it faces significant challenges due to speckle noise and the high dimensionality of image data. This paper proposes an optimized ATR pipeline that integrates a robust multi-stage pre-processing sequence with a novel Two-Stage 2-Dimensional Principal Component Analysis (2DPCA). By preserving the spatial structure of image matrices and performing dual-axis compression, the method achieves a 97.55% recognition rate on the MSTAR benchmark dataset while reducing computational complexity by over 90% compared to traditional PCA methods.
Introduction
SAR images are inherently difficult to process due to multiplicative speckle noise and background clutter. Traditional feature extraction techniques, such as standard Principal Component Analysis (PCA), typically require flattening 2D image matrices into 1D vectors. This "vectorization" process leads to two primary issues: the loss of inherent spatial relationships between pixels and the "curse of dimensionality," where the resulting high-dimensional covariance matrices are computationally expensive and prone to estimation errors.
While standard 2DPCA addresses some of these issues by working directly with matrices, it often only eliminates correlations between rows, leaving feature matrices larger than necessary. This study introduces a two-stage approach to compress both row and column dimensions, significantly enhancing efficiency and accuracy.
Methodology
1. Advanced Pre-processing Pipeline
To isolate the target and mitigate noise, the authors developed a comprehensive pre-processing sequence:
- Logarithmic Transformation: Converts multiplicative speckle noise into additive noise, normalizing the image histogram.
- Adaptive Thresholding & Morphological Filtering: Segments the target from the background and removes "salt and pepper" noise.
- Geometric Clustering: Identifies the largest connected component to isolate the vehicle from residual clutter.
- Power-Law Transformation: Enhances contrast (using an optimal exponent of $\alpha \approx 4$).
- 2D Fast Fourier Transform (2DFFT): Provides translation invariance, ensuring recognition is independent of the target's exact position.
2. Two-Stage 2DPCA Feature Extraction
The core innovation lies in the dual projection method:
- Stage 1 (Row Projection): Projects the image matrix to eliminate row-wise correlations.
- Stage 2 (Column Projection): Transposes the result of the first stage and applies 2DPCA again to eliminate column-wise correlations.
This results in a highly compact feature matrix (reduced from $128 \times 8$ in standard 2DPCA to $12 \times 22$ in the two-stage version). Classification is then performed using a simple Nearest Neighbor Classifier (NNC) based on Euclidean distance.
Experimental Results
The system was evaluated using the MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset, focusing on three vehicle types: BMP2, BTR70, and T72.
Recognition Performance
The proposed method achieved a peak recognition rate of 97.55%. It significantly outperformed several benchmarks, including Template Matching (40.76%), standard SVMs (90.92%), and various Kernel PCA approaches.
Computational Efficiency
By avoiding vectorization, the two-stage 2DPCA reduced the operations required to find features from 696,320 (standard PCA) to just 132,096—a reduction of over 90%. This efficiency makes the system highly suitable for real-time applications and systems with limited storage.
Conclusion
This research demonstrates that preserving the 2D spatial structure of SAR images is vital for accurate target recognition. The combination of rigorous pre-processing and two-stage dimensionality reduction provides a robust, efficient, and highly accurate solution for SAR ATR. Future work may explore the integration of these compact 2D features with deep learning architectures to further improve robustness against varying azimuth angles and smaller training sets.
Summary updated