基于卷积神经网络的SAR图像去噪新方法

A New SAR Image Denoising Method Based on Convolutional Neural Network

本文针对SAR图像相干斑噪声提出一种基于卷积神经网络的SAR图像去噪新方法。把带噪SAR图像输入到本文设计的网络中,利用网络的卷积层提取SAR图像的深层特征,池化层进一步处理以使深层特征降维,再通过反卷积层得到与输入图像尺寸大小一样的结果,并将其与原无噪图像对比,以两者的误差作为神经网络优化器的输入,驱动网络更新各层参数使误差函数最小。为了缩短网络训练收敛的时间,引入ReLU激活函数和批归一化处理,经过少次训练后,网络输出的结果就能接近原始SAR图像。实测数据试验结果表明,与传统SAR-BM3D和SAR-NL去噪方法相比,新方法去噪能力更强,图像视觉效果更好。

In this paper,a new SAR image denoising method based on convolutional neural network is proposed to address the coherent speckle noise problem of SAR images.After an SAR image with noise is input to the designed network,the deep features of the SAR image are extracted by the convolutional layer of the network,followed by further processing by the pooling layer to reduce dimensionality of the deep features,and then an image with the same size as the input image is obtained by using the deconvolution layer.The error between the obtained image and the original image without noise is used as the input of the neural network optimizer to drive the network to update the parameters of each layer,minimizing the error function.In order to shorten the time of network training and convergence,the ReLU activation function and batch normalization are introduced.The network output is close to the original SAR images after a few training sessions.A test is carried out using measured data.The test results show that the proposed method has higher denoising capability and better image visual effect than traditional SAR-BM3D and SAR-NL denoising methods.