研究生
蔣佳欣
指導教授
陳定宏
摘要
以影像內容來進行相片之檢索與分類最近得到廣泛的重視。最常使用的方
法是利用相片之低階影像特徵或是語義註釋來分類。然而這些方法都還不能很
準確的進行相片場景分類。因此若是加上相片中的所內含的EXIF 資訊，將可
提升分類之準確率。EXIF 所包含的的閃光燈（Flash Fired）、焦距（Focal
Length）、曝光時間（Exposure Time）在分辨室內/戶外相片上具有顯著的辨識
力。本篇論文中將整合相片的低階影像特徵值與EXIF 資訊，輔以倒傳遞類神
經網路（Back-propagation neural network）來進行相片室內/戶外分類。
經過第一步驟戶外/室內的分類後，我們將再針對戶外的相片做進一步的辨
別。因為戶外的相片大都包含建築物或是自然風景，而建築物相片相較於自然
風景照，其線條構造大都充滿規則性。因此我們將可利用邊線（edge）來作為
主要的辨別特徵。首先利用canny edge method 找出邊緣（edge），再計算出邊
緣的角度，繪製成邊緣直方圖（edge hitogram）。因建築物大都由垂直與水平
線條構成，所以繪製出的邊緣直方圖，在接近垂直與水平的邊緣角度上，會呈
現出雙峰的特性。因此我們可利用此特性來辨別建築物或是自然風景。
Abstract
The content-base image retrieval and classification has obtained more attention
for the recent application. The most commonly used methods are applying low level
image features or keywords to classify the image contents. However, these methods
are not very accurate. In this thesis, we try to utilize the additional information of
digital pictures to raise the classificatoin accuracy. The EXIF information such as
Flash Fired, Focal Length and Exposure Time, are the embedded information to
record the image information during the picture capture moment. With the additional
EXIF information and the low level image features, we can easily classify the
pictures from the indoor and outdoor scenes. The back propagation neural network
here plays a classifier to recognize the indoor and outdoor scenes.
We will proceed the building/landscape classification from the outdoor scene
pictures. The most signification feature of a building picture is the amount of the
vertical and horizontal edges. First, we use canny edge operator to detect the edges
of the picture. Then we calculate the angle of edge and form a edge histogram. The
experiment shows the edge histogram of a building picture is bi-modal, i.e. the
distribution shape looks like a 「W」. The experimental results shows the proposed
methods can efficiently classify the digital pictures from the indoor/outdoor and
building/landscape scenes.
200) {this.height=200;}" height="200">
200) {this.height=200;}" height="200">