Weighted residual network for SAR automatic target recognition with data augmentation.

Area of Science:

• Radar Systems Engineering
• Artificial Intelligence
• Image Processing

Background:

• Synthetic Aperture Radar (SAR) automatic target recognition (ATR) faces persistent challenges including noise susceptibility, large dataset requirements, and lengthy training times.
• Deep learning has advanced SAR ATR, but existing limitations hinder widespread application and optimal performance.
• Overcoming noise and data scarcity are critical for robust SAR ATR systems.

Purpose of the Study:

• To develop a novel data augmentation technique for SAR images to address noise and data limitations.
• To introduce an efficient network architecture, weighted ResNet, for improved SAR ATR performance.
• To reduce training duration and enhance the accuracy of SAR automatic target recognition.

Main Methods:

• A data augmentation strategy involving controlled addition and removal of speckle noise to expand training datasets.
• Development of a modified ResNet architecture, termed weighted ResNet, incorporating residual strain controls for computational efficiency.
• Utilizing noise perturbation to artificially increase the diversity and scope of training data for SAR ATR models.

Main Results:

• The proposed data augmentation method significantly reduces model training time when combined with the weighted ResNet.
• Experimental analysis confirms improved SAR ATR capabilities using the novel approach.
• The weighted ResNet model demonstrates enhanced performance with reduced data requirements.

Conclusions:

• The integrated approach of data augmentation and weighted ResNet offers a significant advancement in SAR ATR.
• This method achieves superior computational efficiency and recognition accuracy compared to existing SAR ATR techniques.
• The proposed strategy provides a valuable solution for enhancing SAR image analysis and target recognition.