基於色彩味覺之自動化深度調色研究

Abstract

色彩是影像敘事的關鍵要素之一，影視製作的色彩設計在前置階段就已開始；美術指導針對劇情調性做場景的色彩設計，妝髮會做測試，戲服會配合場景及劇情作色彩搭配，攝影指導決定畫面構圖、燈光呈現的方式等，這些都會對整部影視作品的色彩及質感產生影響。殺青後將設計好的色彩交由調光師將所有的色彩及質感設計做最完美的呈現，憑著豐富經驗所累積的美學藝術營造特定情緒氛圍，引導觀眾走入劇情裡。因此，調光師在影視的後期製作裡扮演著極其重要的角色。近年來，隨著OTT（over-the-top）與自媒體的盛行，各播映平台對於影像作品的需求量大增，但後期調色耗時、費工且費用昂貴，因此，若能設計自動化的調色方法必定能解決這些問題。本研究將學術與實務結合，透過深度學習（Deep Learning）的方式，設計一套自動化的一級調色（Primary Color Grading）方法，以味覺中的酸、甜、苦、辣色調風格作為深度學習的目標。在客觀的效果評估方面，峰值信噪比（Peak Signal to Noise Ratio, PSNR）的數值皆在20以上，另外，在結構相似性指標（Structural Similarity Index Measure, SSIM）方面，數值都接近1，足證用此方法能產製出優良的色調影像品質。在主觀評估上，受測者觀看自動化深度調色（Deep Color Grading）的動態影像後，因色彩共感覺（Chromes Thesia）的作用而有高達82%的色彩味覺識別率，皆能指出該色調所要傳達的味道，這意味著此方法所產製出的色調風格受大眾認同。此外，相較於人工調色的動態影像，雖然總體分數深度調色只有32.81%的認同度，但其中苦味深度調色的認同度卻比人工調色高出26%，證明此方法有很大的進步空間與潛能。綜觀之，本研究所設計的自動化深度調色方法具有相當大的可行性，雖然目前無法直接應用於調色作業，但可以為非色彩專業者產出參考圖像。

Color is one of the key elements of image narrative, and the color design of film and television production has begun in the pre-production stage; The art director will do the color design of the scene for the plot tone, the makeup and hair will be tested, the costume will match the scene and the plot for color matching, and the director of photography will decide on the composition of the picture, the way the lighting is presented, etc., which will have an impact on the color and texture of the entire film and television work. After the completion of the design, the designed colors are handed over to the dimmer to make the most perfect presentation of all the colors and texture designs, and the aesthetic art accumulated by rich experience creates a specific emotional atmosphere and guides the audience into the plot. Therefore, the dimmer plays an extremely important role in the post-production of film and television. In recent years, with the prevalence of OTT (over-the-top) and self-media, the demand for video works on various broadcast platforms hasincreased significantly, but the later color grading is time-consuming, labor-intensive and expensive, so if you can design automated color grading methods, you can definitely solve these problems.This study combines academic and practical methods, and through deep learning, designs an automated primary color grading method, taking the sour, sweet, bitter, and spicy color styles in the taste sense as the goal of deep learning. In terms of objective effect evaluation, the peak signal to noise ratio (PSNR) value is above 20, and in addition, in the structural similarity index measure (SSIM), the value is close to 1, which proves that this method can produce excellent tonal image quality. In terms of subjective evaluation, after the subjects watched the dynamic image of the automated deep color grading, they had a color taste recognition rate of up to 82% due to the role of color co-sensing (Chromes Thesia), which can indicate the taste to be conveyed by the hue, which means that the tonal style produced by this method is recognized by the public. In addition, compared with the dynamic image of artificial color grading, although the overall score depth grading is only 32.81% of the recognition, the recognition degree of the bitter depth color correction is 26% higher than that of the artificial color grading, which proves that this method has a lot of room for improvement and potential. In summary, the automated depth grading method designed by the Institute is quite feasible, and although it cannot be directly applied to color grading operations at present, it can produce reference images for non-color professionals.