CAJ | 학술논문

针对原有基于奇异值分解的最大误差可控的高光谱图像压缩(EC-SVD)算法未充分利用图像光谱矢量间冗余的问题，该文将高光谱图像压缩与聚类结合，提出最大误差可控的高光谱图像聚类压缩算法。分析发现，图像的光谱矢量间相似度越高越有利于得到好的最终压缩效果。因此，算法首先使用 K-均值聚类对高光谱图像像元按光谱矢量聚类，以提高同类光谱矢量间的相似度；其次，对每一类像元分别使用 EC-SVD 算法思想压缩以控制最大误差。论文证明了当高光谱图像的像元个数与波段数之比较大，且聚类类数不大于8时，聚类能够提高图像最终压缩比。最后，设计整体压缩实验仿真流程，并对实际高光谱图像进行数值仿真。结果表明，在相同参数条件下，该文算法比 EC-SVD 算法得到的压缩比和信噪比均有提高，最大压缩比提高了10%左右。该文算法能够有效提高EC-SVD算法的图像压缩效果。
침대원유기우기이치분해적최대오차가공적고광보도상압축(EC-SVD)산법미충분이용도상광보시량간용여적문제，해문장고광보도상압축여취류결합，제출최대오차가공적고광보도상취류압축산법。분석발현，도상적광보시량간상사도월고월유리우득도호적최종압축효과。인차，산법수선사용 K-균치취류대고광보도상상원안광보시량취류，이제고동류광보시량간적상사도；기차，대매일류상원분별사용 EC-SVD 산법사상압축이공제최대오차。논문증명료당고광보도상적상원개수여파단수지비교대，차취류류수불대우8시，취류능구제고도상최종압축비。최후，설계정체압축실험방진류정，병대실제고광보도상진행수치방진。결과표명，재상동삼수조건하，해문산법비 EC-SVD 산법득도적압축비화신조비균유제고，최대압축비제고료10%좌우。해문산법능구유효제고EC-SVD산법적도상압축효과。
Aiming at the problem that the maximum Error Controllable compression based on SVD (EC-SVD) algorithm can not make full use of spectral vectors’ redundancy in hyperspectral image, a hyperspectral image compression algorithm with maximum error controlled based on clustering is presented in this paper, by combining hyperspectral image compression with clustering. It is found that a higher compression ratio can be achieved as spectral vectors’ similarity increases. Using the K-means clustering algorithm, the pixels of hyperspectral image are clustered by spectral vectors to improve the similarity of spectral vectors in the same class. Then, the pixels in each class are compressed using the idea of EC-SVD algorithm. And it is shown that the compression ratio increases if the cluster number is no more than 8 and the number of pixels is larger than that of bands in the clustered hyperspectral image. Finally, a total simulation procedure of the improved compression algorithm is designed and some hyperspectral images are tested. The results of the tests show that compression ratios and signal to noise ratios are higher than those of EC-SVD algorithm under the same parameters; the maximum compression ratio rises around 10 percent. The presented improved algorithm can raise the compression efficiencies of hyperspectral images.