基于空间和光谱内容的高光谱图像压缩  

Special and spectral content-based hyperspectral image compres

作　　者：潘少明[1] 顾晓林 种衍文[1] PAN Shaoming;GU Xiaolin;CHONG Yanwen(State Key Laboratory of Information Engineering in Surveying,Mapping,and Remote Sensing,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079

出　　处：《华中科技大学学报（自然科学版）》2023年第9期74-80,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家重点研发计划资助项目(2022YFB3902804);国家自然科学基金资助项目(41671382,62072345).

摘　　要：为提高高光谱图像的压缩性能,提出一种同时利用高光谱图像的光谱信息和空间信息的深度卷积神经网络压缩方法.主要通过主成分分析对高光谱图像进行光谱维降维预处理,在保持图像空间结构特性的同时,去除光谱冗余性.在此基础上,在编码端利用重要性图网络对压缩编码进行内容自适应码率分配,避免低码率下强边缘或小纹理处码率分配不足,从而提高图像压缩重建质量.在高光谱数据上的实验结果表明:该方法在低码率(0.1844)下依然能达到较好的压缩性能,峰值信噪比为27.2099,结构相似度为0.9224.In order to improve the compression performance of hyperspectral images,a deep convolutional neural network compression method using both the rich spectral and spatial information of hyperspectral images was proposed.Firstly,principal component analysis(PCA)considering the spectral content was used for hyperspectral image spectrum dimension reduction preprocessing so as to remove spectral redundancy as well as keep the spatial structure characteristic of the image,then the importance map network was designed to guide adaptive bit rate allocation considering the spatial content at encoder,which can avoid insufficient bit rate allocation at strong edge or small texture under low bit rate,so as to improve the quality of image reconstruction.Experimental results on hyperspectral data show that the proposed method still achieves good compression performance at low bit rate(0.1844),the peak signal-to-noise ratio(PSNR)is 27.2099,and the structure similarity is 0.9224.

关 键 词：高光谱图像 卷积神经网络 光谱信息 空间结构特征 重要性图 自适应码率分配 

分 类 号：P407.8[天文地球—大气科学及气象学]