基於深度學習之影視二級調色研究

Abstract

電影和電視的調色（Color Grading）任務既重要又極複雜。調色涉及美學和技術，需要訓練有素技術人員、耗費大量時間，在情節中提高視覺吸引力，藉改變意象引導觀眾視覺。在這過程中 ，色彩是影像不可或缺的敘述元素，它在觀賞者中扮演著關鍵重要的角色。色彩可突顯影像主體張力，引起人們關注。場景交替、色彩變化都由調光師擔負起重要任務，校正顏色維持藝術價值以取悅人眼，隱藏著色中的不連續性，微妙調整鏡頭。調色，更是一個相當不容易操縱領域。當作業時效性成為商業製片重要考量時，使用自動方式解決是一個受歡迎且省錢選項，所以迅速取得值得參考的深度調色影像，有其高度價值。本研究結合調光與人工智慧跨領域應用，設計以食物顏色、味覺中酸、甜、苦、辣的影像主體二級自動色彩轉換方法。此為食物味覺色調及有關凸顯主體影像二級自動色彩轉換創新嘗試，實際轉換快速且便利。轉換結果依客觀評量之峰值信噪比（PSNR）平均數據為31.29。結構相似性指標（SSIM）平均數據為0.956。從這些數字足以證明此二級自動色彩轉換應用之可實踐性。依主觀評量之（深度調色之判斷酸甜苦辣正確率）平均為61.76％，表示超過六成受測者可以精準分辨深度調色四種味覺。但在接近四項味覺目標色選擇深度調色平均為25%，只有四分之一的專業及非專業人士認為深度調色比人工調色好。綜合以上數據。充分驗證此方法的可行性及實用性。深度調色確實有效逼近人工調色，可以有效節省後期製作時間與費用。雖然深度調色仍有進步空間，但對於未具調光技能與設備的一般使用者而言，具有方便輔助性。

During the television and film post production process, color grading plays an important role. The complex procedure involves delicate technology and theory of aesthetics. Color grading definitely is a time-consuming, semi-art work and has to be practiced by highly trained technicians. Nice color grading work may enhance the visual appeal of plot texture to guide the viewing vision. Color, of course is the image indispensable narrative element that can highlight the main tension of image and attract viewers’ eyeballs. Nevertheless, while correcting the color of filming material to hide some discontinuities, the artistic purpose design must be maintained. For commercial-oriented production considerations, an near-automatic color grading method is a valuable option. Time saving means cost down as well for post production com panies.This research project combines the cross-field application of color grading and artificial intelligence, in order to design a secondary automatic color grading method for image subjects based on good color, good flavors (sourness,sweet,bitterness,spicy hot)。The main purpose is to make the conversion accurate, fast and convenient. I found that :the averages data of the peak signal-to-noise ratio（PSNR） of the conversion result is 31.29 according to the subjective evaluation. The average data of structural similarity index（SSIM） is 0.956.Those comparative data can prove the secondary automatic color grading is effective and feasible.Again, according to the subjective evaluation, the average correct rate of sourness, sweet, bitterness, spicy hot is 62.76%. That means more than 60% members of the 17 persons focus group can accurately differentiate the four flavors after deep color grading. The above-mentioned result was the first part of investigation. I found something interesting in the second part: only 25% of the focus group members who are broadcasting professionals and non-professionals as well considered the deep color grading was better than manual color grading. This data indicated that a lot of users are still approach to manual color grading because of the working mode and sense of unfamiliarity. I wrap up the result: the deep color grading method has the effects close enough to manual color grading. It does save time and cost of post production. Its feasibility and practicability have been verified.Although there are plenty of room for improvement in deep color grading, I can not deny it had a great potential to be polished and be accepted by users in broadcasting industry in the near future.