A Concurrent and Hierarchy Target Learning Architecture for Classification in SAR Application.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Automatic Target Recognition (ATR) is crucial for analyzing Synthetic Aperture Radar (SAR) imagery.
• Convolutional Neural Networks (CNNs) have demonstrated significant success in feature learning for various tasks.
• Extracting robust features from SAR data remains a challenge for improving ATR performance.

Purpose of the Study:

• To propose a novel framework for enhanced Automatic Target Recognition (ATR) in SAR images.
• To investigate the effectiveness of different CNN configurations and feature extraction strategies.
• To improve the accuracy of SAR target classification through advanced feature fusion techniques.

Main Methods:

• Developed three CNN models with varying convolution and pooling kernel sizes.
• Employed two scenarios for image feature generation: fully connected layer activations and Fisher Vectors (FVs) from the last convolutional layer.
• Utilized combination and fusion approaches to integrate features for final classification.

Main Results:

• The proposed framework demonstrated superior performance compared to existing state-of-the-art methods on the MSTAR dataset.
• Experimental results validated the capability of the CNN-based feature extraction and fusion approach.
• The method effectively extracts hierarchical features for improved SAR target recognition.

Conclusions:

• The proposed framework offers a promising approach for enhancing ATR in SAR imagery.
• CNNs, when combined with advanced feature fusion, can significantly improve classification accuracy.
• This research contributes to the advancement of automated analysis of SAR data.