基于优化深度学习的SAR图像地物分类检测方法研究

Study on SAR Image Object Classification and Detection Method Based on Optimizing Deep Learning

针对深度学习在SAR遥感图像地物分类检测中存在的问题,文章通过对基于深度学习的卷积神经网络(Convolutional Neural Network,CNN)进行优化改进,从而提高分类检测准确性。首先提出采用Leaky ReLU函数作为非线性整流函数,克服网络反向传播时梯度消失的问题;然后提出变步长动量梯度下降算法,加速网络收敛、减弱震荡,并避免网络陷入局部极小值。最后综合提出了"Leaky ReLU+变步长+动量梯度下降"的优化方法。通过实验,验证了文章所提出方法的有效性和准确性。

In order to improve performance of convolutional neural network（CNN） algorithm for object classification and detection in SAR image, the parameters of CNN are optimized. Our model uses Leaky ReLU as nonlinear rectification function, the disappearance of gradient has overcome. The gradient descent algorithm with momentum and altered step size is proposed which not only can avoid network falling into local minimum value but also rate and oscillation during convergence of CNN are improved. Finally, comprehensive optimized method is proposed. Experimental results have proved validity and accuracy of above algorithm and method.