應用於數位相機之預先分割感興趣區域的場景分類系統

Abstract

近年來，由於網際網路的普及，數位影像的使用率大幅提昇，帶動數位相機的使用風潮，但是，原本就提供許多功能的數位相機，為了讓使用者在不同場景(scene)下，都能拍出曝光正確的好照片，都會在數位相機上提供不同的場景模式(scene mode) ，例如風景、海灘…等，供使用者選擇，這讓相機在使用上更為複雜且不方便。如果有單一模式能用於拍攝不同的場景，即能解決使用者需時常切換場景模式的不便。 本篇論文提出了預先分割感興趣區域的場景分類系統，稱為PSROI，它可以在相機執行對焦(S1)的同時，將想要拍攝區域的場景進行分類，並在拍照(S2)後，給予相對應的參數設定值，例如光圈、曝光補償值…等。除此之外，本系統還可整合數位相機對焦系統對焦後的結果到場景分類系統中，依照不同的對焦結果，我們可以動態給予場景分類系統不同的權重(weight)，我們稱此權重為對焦權重(focus weight)，使得場景分類出來的結果能更符合使用者所看到的景像(user vision)。在場景分類系統中，我們預設可分類的場景為人像、風景及沙灘/雪景三種。為了更快速達到場景分類的目的，我們減少影像運算區域、以及應用簡單跟較少運算量的演算法來建構我們的系統。實驗結果證明我們提出的系統架構能有效地將一張影像在0.2秒內正確地分到所屬的類別。

With the popularity of internet in recent years, the usage of digital image is growing dramatically, which raises the trend of using digital cameras. In order to allow users taking good photos with correct exposure in difference scenes, the digital camera provides many scene modes such as scenery, beach…etc for user choosing. However, many scene modes complicate the function of digital cameras and users always feel inconvenient to switch the modes again and again. If a single mode can apply in different scenes, it would solve this inconvenience. The thesis proposes Pre-Segmented Region Of Interest Classification Scene System, called PSROI, which is able to classify the taking scenes in the time of digital camera processing focus (S1) and give the parameter such as aperture, exposure compensation…etc after taking photos (S2). Besides, it integrates the Focus system. Following different focus result, the different weight is given in the classification scene system, called Focus weight. Focus weight makes the result of classification scene more suitable for user vision. In the classification scene system, three scenes are set, portrait, scenery, and beach/snow. In order to meet the goal of getting the result of classification more quickly, the operation region of image shrinks and simple and less computation is applied to build the system. The result of experiment proves the proposed system is capable of classifying effectively the scenes into the right categories within 0.2 seconds.