77 / 2024-08-15 02:37:56

Causal Disentanglement Learning for Background Biased SAR Target Recognition

Synthetic Aperture Radar Target Recognition, Causal Disentanglement Learn-ing, Structural Causal Model, Deep Learning.

Synthetic Aperture Radar (SAR) target recognition has important applications in both military and civilian domains, but its performance is limited due to background clutter and target complexity. In this paper, we propose a causal disentanglement learning-based method to improve the robustness of target recognition in SAR images, especially under background clutter conditions. The method aims to reduce the influence of the background on the target recognition by decoupling the target features from the background features to improve the robustness of the recognition system.  And the method provides a new perspective and methodology in the field of SAR target recognition, which is of great significance for improving the performance of target recognition in complex environments. The method consists of the following key steps: first, feature extraction from SAR images using a feature extractor; second, introduction of a causal inference framework that explicitly distinguishes between target and background features through a structural causal model (SCM); third, design of a disentanglement learning strategy that forces the network to learn independent feature representations of the target by designing loss functions with a binary learnable mask and counterfactual samples; and finally, experimental validation of the proposed method on multiple SAR ground target datasets.