Interpretable SAR Target Recognition Based on White Box Neural Network

Synthetic Aperture Radar (SAR) as an all-weather, all-time imaging radar has extensive application value in both military and civilian fields. Although the application of deep learning technology has significantly improved the accuracy of SAR image target recognition, the lack of transparency and interpretability in traditional neural networks greatly limits the application of deep learning models in the SAR field. To address the "black box" issue of traditional neural network models and further enhance the application of deep learning models in the SAR field, this paper introduces the recent interpretable white-box neural networks into SAR target recognition task and discusses the potentials in terms of the performance, robustness and interpretability. The applied white-box neural networks are designed mathematically derivable and implemented with neural network structures. The decision-making process of the model can be analyzed in detail to gain a deeper understanding of its internal mechanisms. Extensive experiments are conducted on the MSTAR dataset. The results show that the white-box neural networks not only achieved high accuracy but also demonstrated strong interpretability and robustness with noisy labels. This research provides new insights into the development of SAR target recognition technology and lays the foundation for the design and implementation of future interpretable and intelligent radar systems.