記錄印刷色彩的計量隨著光學測量儀器的演進，由單一的濃度數值提升至多維度的光譜反射率，這也令印刷色彩估計技術快速發展。然而，既有的理論存在某些假設與限制，使這些理論應用於現今的半色調印刷色彩估計時產生較大的誤差，因此本研究期望可以找出一項符合實際印刷條件的印刷色彩估計方法。
本研究由文獻探討的過程發現非負矩陣分解法（NMF）具有合理分析光譜的特性，因此嘗試將NMF應用於建立一套新的印刷混色估計方法，經過測試後發現NMF可在光譜吸收率與反射率濃度的色彩空間中，有效萃取印刷單色的色彩特徵作為成分，並且成分數量符合實際印刷使用的色墨數量，在二次色以上的色彩特徵萃取雖然僅能分離紙張與色墨的成分，但是可以利用從單色已取得的成分，成功將二次色的色彩特徵分解為單色成分的組成，此方法能透過非負約束的線性迴歸深入分析印刷時色墨與紙張間的關係變化，並且在印刷色彩估計中，此方法估計混色結果的效果皆較使用Neugebauer方程式要佳，此外將迴歸係數以10為底的對數進行轉換，可以得到和網點面積比例百分比良好的多項式預估模式。
The measurement of color in graphic arts application has advanced from a single density value to multi-dimensional spectral reflectance values with the evolution of optical measuring instruments. However, there are some assumptions and limitations in the existing theories, which resulting in large error when they are applied to the half-tone printing color estimation. Consequently, this study aims to find a printing color estimation method that matches with the actual printing conditions. This study finds that the non-negative matrix factorization (NMF) method is suitable for analyzing the spectrum, and attempts to apply NMF method to establish a new set of printing process color estimation method. After testing, it was found that NMF can effectively extract the color characteristics of the primary inks in the color space of spectral absorption and spectral reflectance concentration, and the resulting number of components conforms to the number of color inks used in actual printing. However, the color feature extraction from the overlapping inks can only separate the components of the paper and one type of color ink. The components that have been obtained from the primary inks are then used to successfully decompose the color characteristics of the overlapping inks. The relationship between the original dot areas and the actual ink on paper can be analysed by non-negative constraint linear regression. In printing color estimation, this method results in better color mixing estimating than using Neugebauer equations. In addition, by transforming the logarithm of the regression coefficient to 10 base, one can get a good estimation of the dot areas by polynomial model.
