多尺度字典学习的高光谱图像压缩算法
Hyperspectral image compression using multi-scale dictionary learning
- 2015年19卷第2期 页码:263-272
纸质出版日期: 2015
DOI: 10.11834/jrs.20154069
扫 描 看 全 文
浏览全部资源
扫码关注微信
纸质出版日期: 2015 ,
扫 描 看 全 文
[1]徐大卫,张荣,吴倩.多尺度字典学习的高光谱图像压缩算法[J].遥感学报,2015,19(02):263-272.
XU Dawei, ZHANG Rong, WU Qian. Hyperspectral image compression using multi-scale dictionary learning[J]. Journal of Remote Sensing, 2015,19(2):263-272.
结合小波变换及字典学习提出了一种针对高光谱图像的压缩算法。该算法首先通过小波变换构建多尺度样本集
在小波域使用K-均值奇异值分解(K-SVD)方法学习得到原子尺寸不同的多尺度字典
然后在稀疏表示的过程中
定义一个原子使用频次筛选因子
通过统计局部最优波段稀疏表示时原子使用情况
结合筛选因子对字典原子进行优化筛选
使用精简后的字典对其余波段进行稀疏求解
最后针对不同尺度的表示系数采用自适应的量化编码。实验结果表明
与目前常用的3D-SPIHT和其他的多尺度字典学习算法相比
本文算法在中低比特率下
具有更好的重建性能。
Compression can significantly decrease hyper spectral images to relatively manageable sizes
thereby facilitating their efficient transmission and storage in aground station. Dictionary learning and sparse representation perform well in natural image compression. This study focuses on the properties of hyper spectral images and presents an efficient compression algorithm based on wavelets and dictionary learning for hyper spectral images.First
multi-scale training samples are built through wavelet decomposition; the training samples of each scale are sent to the K-SVD dictionary learning model to obtainmulti-scale dictionaries by joint training
whichsimultaneouslycalculates the errors and update the multi-scale dictionary. Second
statistical analysis is performed on the used dictionary atoms in local optimal bands in the process of sparse coding
and a frequency selection factor is introduced. The statistical information and frequency selection factor are then used to decrease on-used or rarely used atoms in the dictionary. Other bands can be sparsely and easily represented using the simplified dictionary. Finally
the simplified dictionary is directly entropy coded
and the DC component is entropy coded after 4-neighborhood prediction and differential pulse code modulation. The indices of the coefficients of each scale are rearranged according to the numerical value and are separately entropy coded after DPCM. The sparse coefficients are also rearranged according to the sequential changing of indices and are entropy coded together after adaptive quantization. Results show that the proposed scheme outperforms the traditional spatial and other multi-scale dictionary learning algorithms. It is also much better than3D-SPIHT in terms of bit rate distortion performance. As JPEG2000( Part 2) largely benefits from the embedded block coding with optimized truncation strategy
it can achieve a better performance than our scheme. However
our proposed is much faster than JPEG2000( Part 2). This study designed a novel hyper spectral image compressor based on wavelets and dictionary learning.Experimental results reveal that the proposed scheme outperforms the 3D-SPIHT and most compression algorithms. It is also much faster than the state-of-art compression standard JPEG2000( Part 2). This compression algorithm can be improved further in the process of rearrangement before entropy coding. Our experiments prove that dictionary learning and sparse representation theories have great potential in hyper pectral image compression and interpretation applications. This study can motivate future research in this field.
图像压缩高光谱图像多尺度字典学习K-SVD小波变换
image compressionhyper spectral imagemulti-scale dictionary learningK-SVDwavelets
相关作者
相关机构