兼顾色度和光谱精度的多光谱图像LabW2P编解码

针对可见光多光谱图像在通用领域的应用，为提高压缩效率，有效提升重建光谱曲线的色度及光谱精度，进一步存储传输，提出了一种非线性光谱反射率模型，并基于此设计了复杂度适中、光照稳定性好且支持光谱跨设备再现的LabW2P编解码算法。首先根据多光谱图像物理特性，提出非线性光谱反射率模型，将光谱数据表示为线性成分和差别光谱，线性成分由六维变换空间及光谱投影系数组成，差别成分为非线性表示成分，该模型用于光谱数据至不同基变换空间的分解及表示，为算法的构建，光谱及色度重建性能的提升，提供了理论基础；然后，根据人眼视觉系统特征、光照条件，借助CIE标准色度空间转换函数，提取光谱反射率中的三维色度信息Lab，保证重建图像的色度精确性；基于光谱非线性表示模型，采用类视觉曲线的三角函数基，提取线性成分前两维投影系数作为光谱编码的后两维W₁和W₂，用于近似描述CIERGB色度空间中R和G通道，同时有效提高光谱数据的色度和光谱还原度；利用误差补偿机制生成预测差别光谱，采用主成分分析（PCA）法提取其第一维主成分作为编码值P，补偿了线性光谱重建误差，并进一步提升了光谱精确性；最后，组合提取的三部分数据，形成LabW2P编码。LabW2P解码即编码的逆过程。首先，根据Lab及W₁和W₂，结合CIELAB至CIEXYZ色度空间转换函数、光照条件、CIE标准观察者色匹配函数、及三角函数基，采用最小二乘回归，获得变换空间上的重建投影系数，进而重建线性光谱数据；然后，根据P值，采用PCA逆变换，获取重建预测差别数据，最后，结合两部分重建数据，获得光谱重建图像。实验分析显示，LabW2P算法的平均色度精度为0.207 6，较经典的PCA，LabPQR和LabRGB法分别提升了81.54%，55.48%，32.29%，最大平均色差为0.507 0，此外均处于0～0.5之间，达到了视觉难以辨认的可忽略色差的色彩重建水平；平均光谱精度为0.012 7，较PCA性能稍弱，较LabPQR和LabRGB法分别提升了13.01%，6.62%，表明LabW2P编码法的色度和光谱重建性能优势明显。此外该算法可直接用于物体色估计，较PCA和LabPQR法，传输附加信息少，可达压缩比更高。

Multispectral Image LabW2P Codec for Improvement of Both Colorimetric and Spectral Accuracy

In order to improve visible multispectral image (MSI) compression efficiency and further facilitate their storage and transmission for the applications of generic fields, in which both high colorimetric and spectral accuracy are desired, a nonlinear spectral reflectance model is proposed. Then based on this, LabW2P codec is presented, which has the advantages of moderate complexity, good illuminant stability and support for consistent color reproduction across devices. First, according to spectral characteristics of MSIs, a nonlinear spectral reflectance model is proposed for the decomposition and representation of spectral data to different transformation spaces. In the model, spectra are expressed as linear component and difference spectra. The linear component consists of six transformation basis and spectral projection coefficients. And the difference spectra are non-linear represented components. The model provides the theoretical basis for the construction of the coding algorithm and the improvement of spectral and chrominance reconstruction performance. Then, according to the characteristics of human visual system, illumination conditions, and CIE standard chromaticity space transform function, the three-dimensional (3D) colorimetric information Lab of spectral reflectance is extracted to ensure the colorimetric accuracy of the reconstructed image. Meanwhile, based on the nonlinear spectral model, visual-curve-like trigonometric function basis are used to obtain the first 2D projection coefficients of linear component as the latter 2D coding values that are W₁ and W₂, which can be utilized to approximate R and G channels, and also improve the colorimetric and spectral reconstruction accuracy. Combined with error compensation mechanism, predicted difference spectral is generated. The Principal Component Analysis (PCA) method is used to extract the first principal component P which compensates for the linear spectral reconstruction error and further improves the spectral accuracy. Finally, the extracted three components are combined to form LabW2P coding. LabW2P decoding scheme is the inverse of the coding. First, according to Lab, W₁ and W₂, combined with CIELAB to CIEXYZ conversion function, illumination conditions, CIE standard observer color matching function, and trigonometric function basis, the reconstructed projection coefficients on transform space are obtained by using least square regression, and then linear spectral data is rebuilt. Next, based on the value of P, inverse PCA is utilized to gain the reconstructed prediction difference data. Finally, two parts of reconstruction data are combined to get the reconstructed MSI. Experimental results show that the average colorimetric precision of LabW2P algorithm is 0.207 6, which is increased by 81.54%, 55.48% and 32.29% respectively in comparison with that of the classical PCA, LabPQR and LabRGB methods. The maximum average color difference is 0.507 0, and in addition, it is between 0 and 0.5, reaching the color reconstructed level of being visually neglected. Meanwhile, the average spectral precision is 0.012 7, which is slightly weaker than that of PCA, but 13.01% and 6.62% higher than that of LabPQR and LabRGB respectively. The results indicate that LabW2P has obvious advantages of both colorimetric and spectral reconstruction performance. Besides, our coding values can be used directly for object color estimation. And compared with PCA and LabPQR, LabW2P transmits less side information and has higher compression ratio.