RAFnet: SAR Image Autofocusing via Range-Aware Attention and Multi-Scale Loss

Area of Science:

• Remote Sensing
• Signal Processing
• Artificial Intelligence

Background:

• Platform motion errors significantly degrade Synthetic Aperture Radar (SAR) image quality, leading to defocusing and azimuth blurring.
• Accurate autofocusing is crucial for extracting meaningful information from SAR data.

Purpose of the Study:

• To propose a novel autofocusing method for SAR images that addresses platform motion errors.
• To enhance SAR image quality by effectively suppressing phase deviations and sidelobes.

Main Methods:

• Development of a Range-aware Autofocus Network (RAFnet) incorporating a novel range-aware attention module.
• Utilizing 1-D azimuth pooling for feature compression and extraction of high-SNR scattering components.
• Implementation of a progressive multi-scale entropy loss for coarse-to-fine hierarchical learning.

Main Results:

• The proposed RAFnet accurately captures high-level phase fluctuations and effectively suppresses raw phase deviations and sidelobes.
• Quantitative analysis shows a global spatial entropy of 9.8567 and contrast of 5.0091 in focused images when using the attention module and multi-scale loss.
• The method demonstrates superior performance in focusing SAR images affected by motion errors.

Conclusions:

• The Range-aware Autofocus Network (RAFnet) provides an effective solution for high-quality SAR auto-focusing.
• The integration of the range-aware attention module and multi-scale loss significantly improves the accuracy of focus-oriented feature representation.
• This approach offers a robust method for mitigating the impact of platform motion errors on SAR image quality.