Computational interpretability of multilayer preceptron used for SAR image target recognition

The achievement of deep neural networks (DNNs) in the computer vision has aroused great concerns in the synthetic aperture radar (SAR) automatic target recognition (ATR) field. As a simple but effective model, the multilayer perceptron (MLP) is widely used in SAR image target recognition. However, the black-box problem could limit the development of DNNs in SAR ATR. In this paper, we explore the interpretability of MLP from the perspective of computation process of its forward propagation. By using the matrix representation, the function is studied that the angles between parameters and features as well as features magnitudes. Besides, the feature statistics is adopted to discuss the effect of nonlinear activation functions. Finally, some experiments on the MSTAR datasets are carried out and analyzed to demonstrate the effectiveness of the proposed method.