A COMPLEX-VALUED CONVOLUTIONAL NEURAL NETWORK WITH DIFFERENT ACTIVATION FUNCTIONS IN POLARIMETRIC SAR IMAGE CLASSIFICATION

It is well known that the activation function and the gradient descent optimization algorithm have a great influence on convolutional neural network (CNN). Based on the Adam optimization algorithm, this paper proposes a CvAdam optimization algorithm suitable for complex-valued convolutional neural network (CV-CNN), and then in the typical polarization SAR image classification task, Adam and CvAdam were compared using four different activation functions of sigmoid, tanh, Leakey-ReLU and ELU. Experiments on the benchmark dataset of Oberpfaffenhofen show that CvAdam performs better than Adam in both convergence speed and accuracy, no matter which activation function is used.