一套個人化之照片排序系統

由於數位相機的普及造成個人持有的照片張數大幅上升。因此，過去逐張瀏覽挑選喜愛照片的方式已不符合需求。除此之外並非所有使用者均熟悉如何評斷照片好壞的構圖規則，故依需要產生了一連串以照片好壞程度自動做出分類的研究。早期的研究主要多從影像處理的角度出發，以清晰度、亮度、飽和度等低階特徵爲主，但亦能得出令人滿意的結果 (Ke 2006, 72% accuracy, Datta 2006, 86% accuracy)。近來的研究加入了攝影師在拍照時使用的規則(三分法、明暗對比、淺景深)更進一步地提昇了預測的準確率 (Luo 2008, 95% accuracy)。然而過去的研究均使用同樣的研究流程：利用部分資料以機器學習的方式得出一具有預測能力的模型後再用剩下的資料驗證。這樣得出的模型雖然能夠有效的反映出大部分使用者的意見但也許未必能夠適用於單一的使用者。此外先前的研究的正確率大多著重於將照片分爲好與壞兩類而較少著墨於將照片依評分結果做出完整排列之先後順序。在加入使用者意見和探討排序結果之正確率的兩大前提下，我們建立了一套系統可供使用者結合專業的美學分析和個人風格的偏好即時得到重新排序後的結果。本系統可達到(1) 93% 好壞分類準確率 (2) 0.4054的 Kendall Tau 排序相關程度 (3) 92% 的使用者認爲系統重新排序後的結果優於原始排序結果

關鍵字

照片排序；個人化；照片構圖；美學規則

並列摘要

Due to the growing popularity and availability of digital camera, the number of photos owned by each user has increased dramatically. Consequently, manually selecting favorite photos becomes nearly impractical. Moreover, since most digital camera users may not be familiar with the photography rules used by professional photographers, a number of studies on automatic photo selection have been conducted. Most early researches use low-level features (clarity, brightness and saturation) and achieve reasonably good results in binary classification (Ke 2006, 72% accuracy, Datta 2006, 86% accuracy). Recent researches have combined high-level features formulated from photographers’ rules of photo composition with low-level features to produce a better result (Luo 2008, 95% accuracy). However, most of previous researches follow the same framework: training a predication model with some of the data and testing the model with the rest. The problem of this framework is that the prediction model may reflect the preference of a certain group of people but it may not agree with each individual user’s taste. Besides, previous works concentrate on the predication accuracy of binary classification, but we also want to examine the accuracy of ordering of the entire ranked list. Therefore, we create a system for users to combine a ranked photo list based on photography rules with their personal preferences to create more personalized results. Our system has been able to achieve (1) 93 % binary classification accuracy (2) 0.4054 Kendall Tau correlation rank (3) 92% satisfaction rate of personalized re-ranked list over original list.

並列關鍵字

Photo ranking ； Personalized ranking ； Photo composition ； Aesthetic Rules.

參考文獻

[1] W. H. Kruskal, Ordinal measures of association. Journal of the American Statistical Association, 53(284): 814-861, 1958.

[2] R. O. Duda, , P. E. Hart, Use of the hough transformation to detect lines and curves in pictures. Communications of the ACM 15, 1 (1972), 11–15.

[3] R. Chellappa, “Two-dimensional discrete Gaussian Markov random field models for image processing,” Pattern Recognition, Vol. 2, 1985, pp.79-112.

[4] J. Canny, A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 8, 6 (1986), 679–698.

[5] B. S. Manjunath and W. Y. Ma, “Texture Features for Browsing and Retrieval of Image Data,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 18, No. 8, August 1996.

國際替代計量

一套個人化之照片排序系統

全文下載

主題瀏覽