Interpretability of Deep Network Structures for SAR Image Target Recognition Based on Generative Adversarial Mask
Yuhang Zhang, Jie Geng, Wen Rong Jiang · 2023
Deep neural networks (DNN) have been widely used in SAR image Target recognition and have achieved great success. However, the black box attribute of DNN makes the mechanism of the network inexplicable and confusing. In this paper, a network layer analysis model based on generative adversarial mask (GAM) is proposed to enhance the structure interpretability of DNNs. Inspired by network adaptive search, we introduce a mask layer and construct a layer analysis model. After sparse training and generative adversarial training, the contribution of the network layer can be calculated based on the mask layer parameters. We conducted sufficient experiments on several typical DNNs to verify the effectiveness of GAM.