基於支持向量回歸技術之人臉照片審美分類與美化

Digital Face Classification and Beautification Based on Support Vector Regression

劉恒溫(Heng-Wen Liu)

指導教授 ： 歐陽明

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2008年)

https://doi.org/10.6342/NTU.2008.01026

摘要

對相片的修改，可以讓相片中的人變的更美麗，更吸引人，這早就不是什麼祕密了。對專業的攝影師及廣告設計師來說，怎麼讓相片更吸引人更是一門藝術。然而對於沒有專業技巧的我們，即使我們也想要把自己的相片變漂亮，卻也總是力不從心。
我們這一篇論文，就是要幫妳把妳的相片變美麗。不需要有什麼利害的技巧，只要給我們系統妳的相片，我們就幫妳把相片變美麗！我們利用兩百張人臉正面的相片，以及人對於這兩百張相片的評分，使用支持向量回歸(Support Vector Regression)技術做為基礎，讓電腦也能學習人類的審美觀念，對人臉的相片做出評分。在這一個部分，我們的評分系統與人的平均評分比較，取得了相關系數0.64的相關性(35個評分者參與實驗)，相較於人與人平均評分的相關係數的平均為0.68，我們系統的評分結果已與人類評分者的評分結果相近。
有了評分系統後，我們進一步利用貪婪演算法和我們的評分系統，來對照片做美化。方法是微小的改變原來照片的36個特徵點（依照經驗先做分群），以在我們的評分機器上取得最好的分數。得到了分數更高的特徵點後，再將原來的照片變形到新的位置，來取得美化後的相片。經過我們的美化，平均每張相片可以增加1.37分(分數範圍為1到7分)。

關鍵字

人臉 ； 人臉美化 ； 人臉評分 ； 人臉變形 ； 支持向量回歸

並列摘要

It’s not a secret that we can make people in the photo more beautiful and attractive by using application software for modifying procedures. How to make photos looks more attractive is even an artistic issue for professional photographers and commercial designers, however, such modifying procedure is not a simple job for ordinary people.
The goal of this thesis is to make people’s photos look more beautiful. Without any special skills, as long as you give our system your photos, we will help you to get your beautified photos. To reach this goal, our system was divided into two parts, one is the Rating System, and another is the Beautifying System.
The rating system can evaluate a full-face photo, just like what we human usually do. We use the Support Vector Regression (SVR) to train an evaluation model, based on 213 full-face photos, and another 35 people to evaluate them. With this model our system can rate photos like a human, since our rating system achieves a correlation of 0.64 compare to human rating, which is comparable to the average human V.S. human rating of 0.68.
With the previous scoring system, we further use a greedy algorithm to beautify photos. We slightly modify 36 feature points (grouped by heuristics) from the source to the target and try to get better rating by hill climbing. With the higher rating feature points, we can simply warp the source image to the resulting target image. After our beautifying procedure, we can in average increase 1.37 rating points in our rating system (score ranges from 1 to 7 with 7 the best score).

並列關鍵字

face ； face beautification ； face rating ； face warp ； support vector regression

參考文獻

[1] Eisenthal, Y., Dror, G., and Ruppin, E. Facial Attractiveness: Beauty and the Machine. In Neural Comput. 18:119-142, 2005.

[3] Beier, T., and Neely, E. Feature-based Image Metamorphosis, In SIGGRAPH, 26(2):35–42, 1992.

[6] Bookstein, F. L. Principal warps: Thin-plate splines and the decomposition of deformations. In IEEE PAMI, 11(6):567–585, 1989

[7] Cootes, T., Edwards, G., AND Taylor, C. Active appearance model. In IEEE PAMI, 23(6):681-685, 2001

[8] Schaefer, S., McPhail, T., and Warren, J. Image Deformation Using Moving Least Squares, In SIGGRAPH, 25(3):533–540, 2006