融合空谱-梯度特征的深度高光谱图像去噪  

Deep hyperspectral image denoising by fusing space spectrum-gradient features

作　　者：李忠伟[1] 张浩[1] 王雷全[1] 任广波[1] 崔行帅 LI Zhongwei;ZHANG Hao;WANG Leiquan;REN Guangbo;CUI Xingshuai(College of Oceanography and Space Informatics,China University of Petroleum,Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266580

出　　处：《光学精密工程》2022年第5期615-629,共15页Optics and Precision Engineering

基　　金：国家自然科学基金联合基金资助项目(No.U1906217);国家自然科学基金面上项目(No.62071491);中央高校基本科研业务费专项资金资助项目(No.19CX05003A-11)。

摘　　要：为了去除高光谱图像采集过程中产生的噪声,提升后续图像处理的性能,提出了一种融合空谱-梯度特征的深度高光谱图像去噪方法。它包括空谱特征网络和梯度特征网络,且各网络使用密集跳跃连接和可分离卷积策略进行优化。空谱网络模型实现噪声特征的精确提取,梯度网络模型对噪声纹理特征进行补充,最后基于两个网络的特征提取结果进行融合,实现噪声特征的精准刻画,并用于恢复干净图像。分别在合成噪声图像和真实噪声图像上验证方法的有效性。实验结果表明,该方法在恢复图像内部结构上效果显著,在噪声标准差50的条件下去噪结果的平均信噪比达到29.426 dB,平均结构相似性达到0.9678 dB,去噪结果使用支持向量机算法进行分类,分类精度达到90.89%。To remove the noise generated during the process of hyperspectral image acquisition and to improve the performance of subsequent image processing,a deep hyperspectral image denoising method was proposed based on the fusion of spacial spectral and gradient features.It included spacial spectral and gradient characteristic networks,and each network was optimized with a dense connection and separable convolution strategy.The spacial spectral network model extracted the noise features,and the gradient network model supplemented the texture features of the noise.Finally,the feature extraction results of the two networks were fused to achieve characterization of the noise features and to restore clean images.In this study,the effectiveness of the proposed method was verified on synthetic-noise and real-noise images.Experimental results showed that the method had a significant effect on the restoration of the internal structure of images.Under the condition of noise standard deviation of 50,the mean PSNR reached 29.426 dB,while the mean SSIM reached 0.9678 dB.The denoising results and the original image were classified by SVM algorithm,and the classification accuracy reached 90.89%.

关 键 词：高光谱图像 去噪 空谱网络 梯度网络 密集连接 可分离卷积 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]