计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (14): 176-182.DOI: 10.3778/j.issn.1002-8331.1911-0060

• 图形图像处理 • 上一篇    下一篇

改进卷积神经网络SAR图像去噪算法

钱满,张向阳,李仁昌   

  1. 南昌航空大学 信息工程学院,南昌 330063
  • 出版日期:2020-07-15 发布日期:2020-07-14

Improved Convolutional Neural Network for SAR Image Despeckling Algorithm

QIAN Man, ZHANG Xiangyang, LI Renchang   

  1. College of Information and Engineering, Nanchang HangKong University, Nanchang 330063, China
  • Online:2020-07-15 Published:2020-07-14

摘要:

合成孔径雷达(SAR)通常会被一种称为散斑的乘性噪声干扰,这使得图像的解释变得困难。为解决这一问题,提出一种改进卷积神经网络SAR图像去噪方法。对图像进行下采样再对下采样子图像进行卷积提取特征,这可以有效扩大感受野提高去噪效率;为了减少梯度消失问题和提高模型去噪性能,网络又引入了跳跃连接和残差学习策略;利用仿真和实测数据对网络进行测试与评估,实验结果表明提出的方法具有良好的去噪效果和较高的计算效率,对比其他去噪方法,该方法不仅去噪效果好,而且效率更高。

关键词: 合成孔径雷达(SAR)图像去噪, 卷积神经网络, 图像下采样, 跳跃连接, 残差学习

Abstract:

Synthetic Aperture Radar(SAR) is often suffered from a multiplicative noise commonly referred to as speckle which makes the interpretation of images difficult. To remove the speckle noise of SAR images, this paper proposes an improved convolutional neural networks approach for SAR image despeckling. The method firstly downsamples the image and then performs convolution to extract feature of the downsampled sub-image, which can effectively expand the receptive field and improve the denoising efficiency. In addition, skip connections and residual learning strategy are added to the despeckling model to reduce the vanishing gradient problem and improve the performance. Finally, simulated and real SAR images are utilized to test and evaluate the network. Experimental results show that compared with the start-of-art techniques, the proposed method achieves better performance and high efficiency.

Key words: Synthetic Aperture Radar(SAR) image despeckling, convolutional neural networks, image downsampling, skip connections, residual learning