A Robust Causal Diffusion Model for Synthetic Aperture Radar Automatic Target Recognition with Noisy Labels

Jiaxiang Liu, Zhunga Liu, Zuowei Zhang, Shuyue Wang, Xuemeng Hui

Abstract: Based on a large-scale annotated dataset, the deep network models have achieved outstanding performance in synthetic aperture radar (SAR) automatic target recognition (ATR). Due to the interference encountered during measurement process and sample collection, it is inevitable that noise samples marked incorrectly will appear in the annotations, leading to a decline in the performance of the recognition model. In this paper, we propose a new causal diffusion model to eliminate this impact. The core idea is to construct clean pseudo-labels through causal intervention to participate in the training of the diffusion model. Firstly, a structural causal model (SCM) is constructed to describe and analyze the bias introduced into ATR by noisy labels, particularly those stemming from the measurement process and annotation ambiguity. This model, combined with the diffusion model, models the causal paths between variables under the condition of noisy labels from the perspective of a generative model. Secondly, a method called Causal Clean Pseudo-Label (CCPL) is proposed to filter and reconstruct timestep-dependent pseudo-labels from the noisy dataset. The new pseudo-labels can avoid the misleading effects of noisy labels on the diffusion model. Finally, the trained diffusion model is sampled to achieve efficient and robust recognition under noisy labels. Experiments conducted on two SAR datasets also demonstrate that this method can effectively handle the training of networks with varying proportions of noisy labels and outperforms existing state-of-the-art (SOTA) methods.

External IDs:doi:10.1109/tim.2025.3635309