Attribute Feature Perturbation-Based Augmentation of SAR Target Data.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep learning for target detection requires large, diverse datasets.
• Existing synthetic aperture radar (SAR) image augmentation methods suffer from instability, poor quality, and lack of physical interpretability.

Purpose of the Study:

• To propose a feature-level data augmentation method for SAR images to address limitations of current techniques.
• To improve the generalization of deep learning-based target detection and recognition algorithms.

Main Methods:

• Utilized an enhanced capsule neural network (CapsNet) for feature extraction and attribute decoupling.
• Employed an attention mechanism-based framework for effective feature representation.
• Perturbed decoupled features (amplitude, angles, shape) and reconstructed images using pixel and perceptual loss.

Main Results:

• Achieved a peak signal-to-noise ratio (PSNR) of 21.6845.
• Obtained a radiometric resolution (RL) of 3.7114 and dynamic range (DR) of 24.0654.
• Demonstrated superior performance in augmenting SAR target images compared to random noise methods.

Conclusions:

• The proposed feature-level augmentation method enhances SAR image quality and diversity.
• This approach effectively reduces the need for extensive training data.
• The method shows significant improvements for deep learning-based SAR target detection and recognition.