| Abstract: | 影片數目眾多而龐雜，對於一部未知的影片，觀眾很難知道此影片的類型，造成觀眾想要選擇一部想要觀看類型的影片，必須花費很多時間觀看預告片或跳躍式播放該段電影，才可以了解該影片是否屬於自己想要觀賞的類型。本文的目的是要讓觀眾無須先行觀看預告片或影片內容就可知曉影片類型，便可以決定是否屬於自己想看的影片類型；我們透過色調分析，作為影片分類的依據；首先將影片以取樣方式擷取影格，再將影格中色調分類為紅色、綠色、藍色及黃色等四類，並且將電影預告片之HSV(Hue, Saturation, light Value)色彩空間(色相、飽和度及亮度)的標準差、平均值、比例做為影片分類的特徵參數，接著使用線性激活函數之深度學習神經網路將影片分類成文藝愛情類、科幻類、動作冒險類、恐怖驚悚類、喜劇等五大類；為了評估分類的效能，我們使用精確率、召回率、F測度(F-measure)檢測分類的正確性。經由實驗結果證明使用影格色調的數量、連續變化性作為特徵參數，將影片分類成文藝愛情類、科幻類、動作冒險類、恐怖驚悚類、喜劇等五大類確實可行；其中，文藝愛情類及喜劇類以紅色及黃色系居多、科幻類及恐怖驚悚類綠色及藍色系居多、動作冒險類東洋片以黃色系及紅色系居多，西洋片則以藍色系及綠色系居多。透過HSV色調分佈特性，使用規則導向式方法分類影片之後，得到的精確度為83.33%、召回率為100%、F測度為90.71%；當門檻值較為寬鬆時，一部影片可能會有多個分類結果。我們也利用深度學習神經網路進行影片分類，精確率為93.3%、召回率為92% 、F測度為91.4%，分類效果優良。
The number of films is numerous and complex. A viewer wants to choose a favorite movie is time consuming. This thesis aims to develop automatic film classification systems. The proposed methods allow viewers to determine the type of films they want to watch without first viewing the trailer or the video content. Firstly, a film is sampled frame by frame. The color space including red, green, blue, yellow, hue, saturation, and brightness value (HSV) in each frame is analyzed. The mean and deviation of the HSV are computed and utilized as classification features for each film. These features are employed in the proposed rule-oriented classification method and also fed into deep learning neural networks. The trailers are classified into five categories, including science fiction, literature-love, action, comedy films, and horror thrillers. In order to evaluate the effectiveness of the classification, we use the precision rate, recall rate, and F-measure as objective measures. Experimental results show that literature-love and comedy films are mostly red and yellow, science fiction films and horror thrillers are mostly green and blue, and action and adventure Dong-yang films are mostly yellow and red, while western films are mostly blue and green. By using the rule-oriented classification method, the accuracy, recall rates, and the F measure are 83.33%, 100%, and 90.71%, respectively. We also use deep learning neural networks for film classification, with a precision rate of 93.3%, a recall rate of 92%, and an F-measure of 91.4%. The performance is satisfied. |
