Generalizing SAR Object Detection: A Unified Framework for Cross-Source Scenarios

The significant domain differences between heterogeneous synthetic aperture radar (SAR) data lead to a serious degradation of the generalization ability of existing detection models in real cross-source scenarios. Traditional domain adaptation methods rely on target domain data and are difficult to cope with multisource dynamic interference, while existing domain generalization (DG) schemes have not yet adequately addressed the multidimensional domain shift coupling problem unique to SAR images. This letter presents a unified framework for cross-scenery DG heterogeneous SAR image object detection tasks. First, we match the SAR image scattering features with dynamic statistics to expand the domain feature space. Next, domain-invariant features and domain-specific features disentanglement module is designed for the above feature space, and redundant domain discriminative information is removed. In this way, the performance enhancement of both for the DG algorithm can be maximized. Cross-source scenario object detection results across multiple SAR/remote sensing image datasets obtained in a detection head with Faster region convolutional neural network (RCNN) as the baseline network all achieve the state-of-the-art (SOTA) performance available.