[Paper Review] Deep Learning for Target Classification from SAR Imagery: Data Augmentation and Translation Invariance

Hidetoshi Furukawa|arXiv (Cornell University)|Aug 25, 2017

Advanced SAR Imaging Techniques53 citations

TL;DR

The paper analyzes translation invariance of CNNs for SAR ATR and demonstrates that domain-specific data augmentation is key; augmented training achieves 99.6% on MSTAR center-cropped test data.

ABSTRACT

This report deals with translation invariance of convolutional neural networks (CNNs) for automatic target recognition (ATR) from synthetic aperture radar (SAR) imagery. In particular, the translation invariance of CNNs for SAR ATR represents the robustness against misalignment of target chips extracted from SAR images. To understand the translation invariance of the CNNs, we trained CNNs which classify the target chips from the MSTAR into the ten classes under the condition of with and without data augmentation, and then visualized the translation invariance of the CNNs. According to our results, even if we use a deep residual network, the translation invariance of the CNN without data augmentation using the aligned images such as the MSTAR target chips is not so large. A more important factor of translation invariance is the use of augmented training data. Furthermore, our CNN using augmented training data achieved a state-of-the-art classification accuracy of 99.6%. These results show an importance of domain-specific data augmentation.

Motivation & Objective

• Assess how translation misalignment affects CNN-based SAR ATR performance.
• Evaluate the impact of data augmentation on translation invariance and classification accuracy.
• Visualize translation invariance using accuracy-translation maps.
• Compare CNN performance with and without augmentation on MSTAR target chips.

Proposed method

• Use a deep residual CNN (ResNet-18) to classify 96x96 SAR target chips into 10 classes.
• Train models with and without data augmentation (random crops from 100x100 chips) and evaluate on center-cropped 96x96 test patches.
• Define and visualize an accuracy-translation map that shows classification accuracy under x/y translations Δx, Δy.
• Analyze confusion matrices for non-augmented and augmented training to quantify per-class performance.
• Report overall accuracy on center-cropped test data for both settings (non-augmented vs augmented).
• Discuss how augmentation influences translation invariance more than architectural changes alone.

Experimental results

Research questions

• RQ1 What is the relation between data augmentation and translation invariance in CNN-based SAR ATR?
• RQ2 How does augmentation affect overall classification accuracy and per-class performance on MSTAR?
• RQ3 Can translation-induced misalignment be mitigated by training data perturbations rather than network depth alone?
• RQ4 To what extent does ResNet-18 achieve translation invariance without augmentation on aligned SAR target chips?

Key findings

• CNN without data augmentation on aligned MSTAR chips yields limited translation invariance.
• Data augmentation (random crops during training) significantly improves translation robustness and overall accuracy.
• CNN with augmentation achieves 99.6% accuracy on center-cropped test data, reaching state-of-the-art performance in this study.
• Accuracy-translation maps show robust performance within small translations (Δx, Δy within ±2 pixels) when augmentation is used.
• Per-class confusion reduces notably with augmentation, improving accuracies across most target classes.