近來影像分類的研究,大部份是針對特定的應用領域訂定全域式的影像特徵,因此不易應用於不同領域的影像主題,且全域式的影像特徵較不能顯示出個別物件與空間位置的影像特徵。本論文提出以色塊屬性關聯規則作為影像特徵,我們首先將影像轉換至HSV色空間上並量化,以色塊指標串列輔助取出色塊及色塊屬性的計算,接著再建構二元關聯次數累計表以快速計算出一張影像中的色塊屬性關聯規則之支持度與確信度。本論文提出動態多決策樹演算法來挑選出對區分影像類別具重要性的色塊屬性關聯規則,進而建立影像分類決策規則;亦提出影像分類決策規則的精簡方法,可有效降低分類規則的數量,且不明顯影響分類正確率。我們亦將分類方法擴增為模糊型式,可接受具模糊類別的訓練影像並產生模糊分類結果。在我們的實驗中顯示,本論文所提出之影像分類方法分類的正確率明顯優於C4.5與模糊決策樹,並對於各種不同種類的影像領域皆能達到一定程度的分類正確率。
Most previous works on image classification which are purposed for specific image domain, extract global image properties to be feature of an image. However, the global image properties can't represent objects and spatial features well. In this thesis, a kind of object-based image feature is designed, called Block Attribute Association Rules (BAAR), which indicates the relationship among locations and sizes of color blocks. First, the color domain of an image is transformed to HSV color space and quantized to be 148 colors. After that, color blocks and their content attributes are extracted efficiently by applying Block List. The Binary Relationship Counting Table (BRCT) is designed for computing the supports and confidences of BAARs efficently. Moreover, Dynamic Multi-Decision Tree (DMDT) algorithm is proposed for deriving classification rules, and a pruning algorithm is provided to reduce the number of classification rules. The proposed strateies are also extended to perform fuzzy classification. According to the experiment results, it shows that the classification accuracy of proposed classification methods is superior than C4.5 and fuzzy decision tree, and the proposed strateies are applicable on various image domains well.