以ISRF內插法應用於物體頻譜反射率重建之研究

Abstract

近年來，隨著人們對於數位影像品質的要求提升，色彩的準確與否成為一項熱門且重要的研究議題。色彩的形成被光譜分布、物體頻譜反射率與觀測者配色函數三項要素所影響。在實用上，若是能直接由組成數位影像像素中的RGB色頻數值，估計出物體的頻譜反射率，即可透過多頻譜成像方式為原始影像提供更豐富的應用。 本研究重點在提出一個新的物體頻譜反射率重建方法，將藉由真實量測的物體頻譜反射率資料，使用自然鄰點內插法(Natural Neighbor Interpolation，NNI)與ISRF (Idea Spectral Reflectance Family)來估計待測物體的頻譜反射率。 經由實驗數據證實本研究所估計的物體頻譜反射率相當精準，在物體頻譜重建方面，均方根誤差(Root Mean Square Error，RMSE)平均小於0.0747；曲線配適係數(Goodness-of-Fit Coefficient，GFC)平均大於0.9。在色彩顯像模擬方面，在標準照明體D65下，以色差公式∆E 2000評估計算，平均值小於1.6227，在NBS (National Bureau of Standards)標準中皆屬於人眼可忽略的差異程度。同時，另外一項重要的發現即是導入本研究開發之ISRF方法，與其他使用內插重建物體頻譜反射率方法的研究相比，能夠解決顏色落於該模型凸包範圍(Convex Hull)之外而無法被計算的問題。In recent years, with the fast growing demand of the quality of digital image, color accuracy becomes a popular and important topic in the research. Color performance is made from three elements, they are spectral power distributions (SPD), spectral reflectance, and human eyes color matching function (CMF). If we can estimate the spectral reflectance of the objects directly from RGB channel values of the pixel in the image, we can provide more rich applications for the original image by using multi-spectral analysis. In this study, we proposed a new method to reconstruct the spectral reflectance of object by NNI (Natural Neighbor Interpolation) taking limited set of the reflectance data from real measurement, corrected with ISRF (Idea Spectral Reflectance Family) to estimate spectral reflectance of the objects. The experimental results showed the accuracy of our method is high in estimating the spectral reflectance of objects. In the evaluation of spectral reflectance reconstruction, the RMSE (Root Mean Square Error) is less than 0.0747 and GFC (Goodness-of-fit Coefficient) is larger than 0.9 in average. In evaluation of color imaging simulation, under standard illuminant D65, the color difference is less than 1.6227 in average when evaluated by color difference formula ∆E 2000. It reveals that human eyes can not distinguish the difference according to NBS (National Bureau of Standard). Simultaneously, another important finding is that importing our ISRF method can solve the problem of colors outside the convex hull of model which can not be computed when comparing the other methods of reconstruction by interpolation.