Research map / Label-efficient and robust learning / Recognition and long-tail
Gamma Distribution PCA-Enhanced Feature Learning for Angle-Robust SAR Target Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ΓPCA, a Gamma-distribution PCA that derives parameter-free, angle-robust convolution kernels prepended to ResNet101 and ViT-B/16 to improve SAR target recognition under limited azimuth and depression variation on MSTAR.
Problem
Deep SAR target recognizers are sensitive to azimuth and depression angle changes, and training data often lacks coverage of all observation angles, hurting robustness.
Approach
Extends PCA's squared-error objective to the Gamma (exponential-family) distribution via a generalized PCA with a canonical link, yielding a projection matrix used as fixed convolution kernels that capture angle-invariant low-rank features and add no trainable parameters.
Key results
- ΓPCA-ViT reaches 76.02% overall accuracy on MSTAR with training on azimuth 0°-90° at depression 17°, versus 72.45% for ViT-B/16.
- ΓPCA-ViT reaches 75.22% overall accuracy when trained on azimuth 180°-270°, versus 73.19% for ViT-B/16.
- ΓPCA-ResNet improves over ResNet101 by 3.61% and 2.98% on the two training quadrants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ΓPCA to ViT-B/16 | ViT-B/16 · MSTAR overall accuracy; train azimuth 0°-90° | accuracy: 72.45 -> 76.02 | Helped |
| Add ΓPCA to ResNet101 | ResNet101 · MSTAR overall accuracy; train azimuth 0°-90° | accuracy: 63.62 -> 67.23 | Helped |
| Add ΓPCA to ViT-B/16 | ViT-B/16 · MSTAR overall accuracy; train azimuth 180°-270° | accuracy: 73.19 -> 75.22 | Helped |
| Add ΓPCA to ResNet101 | ResNet101 · MSTAR overall accuracy; train azimuth 180°-270° | accuracy: 66.74 -> 69.72 | Helped |
| Add ΓPCA to ResNet101 | ResNet101 · SAR-AIRcraft-1.0 test | accuracy: 97.33 -> 98.67 | Helped |
| Add ΓPCA to ViT-B/16 | ViT-B/16 · SAR-AIRcraft-1.0 test | accuracy: 97.05 -> 98.12 | Helped |
| Add ΓPCA to Swin-B | Swin-B · SAR-AIRcraft-1.0 test | accuracy: 97.41 -> 98.32 | Helped |
| Replace ΓPCA with a conventional convolution layer (ablation) | ViT-B/16 · MSTAR A-R Test; kernel size 5 | accuracy: 76.58 -> 69.69 | Hurt |
| Replace ΓPCA with a wavelet convolution layer (ablation) | ViT-B/16 · MSTAR overall accuracy; kernel size 17 | accuracy: 76.01 -> 70.47 | Hurt |
| Replace ΓPCA with standard PCA (ablation) | ViT-B/16 · MSTAR AD-R Test; kernel size 17 | accuracy: 67.59 -> 63.64 | Hurt |
| Use kernel size 5 instead of kernel size 17 (ablation) | ΓPCA-ViT · MSTAR overall accuracy | accuracy: 76.01 -> 72.56 | Hurt |
Surprising findings
- Standard convolution layers added before ViT often degrade accuracy, while PCA and ΓPCA kernels improve it as kernel size grows.
Limitations
- Evaluated mainly on MSTAR, a single small SAR benchmark.
- On the AD-R test, ΓPCA-ResNet is below plain ViT-B/16 and some baselines, and ΓPCA-ViT is not leading on AD-R in Table 1.
- SAR-AIRcraft-1.0 experiment lacks angle annotations, so it does not test angle robustness.
Experimental setup
Datasets: MSTAR SAR-AIRcraft-1.0
Benchmarks: MSTAR SAR-AIRcraft-1.0
Metrics: overall accuracy test accuracy on azimuth-robustness and azimuth-depression-r
Baselines: ResNet101 EfficientNet-B3 ViT-B/16 Swin-B VisionLSTM-T VisionMamba-T MSNet-PIHA WTConvNeXt-T
Builds on: PCANet generalized PCA (Landgraf & Lee, 2020) ResNet Vision Transformer
Closest papers by meaning
- RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · similarity 0.83)
- Statistically Principled Deep Learning for SAR Image Segmentation (AAAI 2024 · similarity 0.81)
- SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (NeurIPS 2024 · similarity 0.80)
- Unleashing Channel Potential: Space-Frequency Selection Convolution for SAR Object Detection (CVPR 2024 · similarity 0.79)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · similarity 0.79)
- Superposed Atomic Representation for Robust High-Dimensional Data Recovery of Multiple Low-Dimensional Structures (AAAI 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.