利用資料探勘方法辨識物件

Abstract

物件辨識的目的是識別圖片中的物件並且將物件歸類到具有相同特性的類別中，本篇論文中我們利用資料探勘方法找出特徵樣式達成物件辨識的目的，方法主要分成三個階段。第一，從物件的輪廓中找出高曲率點並且將兩點間的距離以及該點的曲率型成可以表示物件的序列；第二階段，利用Apriori演算法找出最大頻繁樣式(maximal frequent pattern)作為各種類別物件的特徵樣式(feature pattern)，這個階段引用模糊觀念因此可以接受物件少許的變化；最後利用近似序列比對演算法(approximate sequential matching)比對物件，這個演算法屬於動態規劃問題，如果特徵樣式中的項目不完全符合於序列中且不符合的數量不大於使用者事前定義的門檻值，這個特徵樣式就符合該條序列，因此該條序列所代表的物件應該被歸類到該特徵樣式代表的類別。 實驗結果顯示，我們提出的方法效果比使用地標的方法好而且不需要很多訓練資料就可以辨識出具有相似姿勢的物件。

The goal of object recognition is to identify the object in an image and objects are classified into several categories which shares a common feature. In this thesis, we proposed a data mining approach consisting of three phases to realize object recognition by a means of mining feature patterns. First, high curvature points are extracted from the contour of an object and the distances between two points and the curvatures at these points are recoded to form a sequence to represent the shape of this object. In the second phase, Apriori algorithm is used to discover the maximal frequent patterns as feature patterns of every category. A fuzzy concept is included into the process of mining feature patterns to tolerance slight transformations. Finally, matching is done by approximate sequential matching which is a dynamic programming problem. If the number of mismatch items between a feature pattern and a sequence is no more than a user-defined threshold, we say the feature pattern is matched with the sequence. Thus, the object of this sequence is classified into the category of this feature pattern. Experimental results show that our method performs better than a landmark-based method and does not require many training dataset to identify objects with similar poses.