SAR Image Ground Object Recognition Detection Method based on Optimized and Improved CNN

摘要

In order to solve overfitting of Convolutional Neural Network (CNN) in Synthetic Aperture Radar (SAR) image ground object recognition detection, in this paper CNN was optimized and improved for obtaining improved performance. Firstly, tanh function was selected as activation function to make CNN stably converge. Secondly, dropout was adopted in hidden layer of CNN to reduce overfitting. For the problem of CNN fell into local minimum value after adopting dropout, this paper proposed a weight update method of the `variable step size + gradient descent with momentum', which could make CNN escape local minimum value and weaken oscillation. Finally, optimized and improved method of the `tanh + dropout + variable step size + gradient descent with momentum' was proposed, which effectively reduced overfitting. Effectiveness and accuracy of proposed method are verified by experiments.