基於兩階段模板匹配演算法的圖形識別

Abstract

在電腦視覺的領域裡，圖形識別是一個重要的議題。圖形識別的困難處，在於如何整合圖形特徵並計算圖形之間的相似度，使其結果能夠成功的鑑別出目標物件。模板匹配法是一種圖形相似度的計算方法，根據比對方式的不同，可分成硬式模板匹配與可變模型匹配兩大類。硬式模板匹配速度快，但辨識準確率不如可變模型匹配，因此本論文結合兩者的優點，提出一個兩階段的模板匹配演算法。我們首先利用硬式模板匹配，找出圖片中可能與目標物件相似的候選圖片，接下來使用可以自由調整彈性的可變模型匹配，對候選圖片進行相似度的計算與排序，達到圖形識別的目的。本論文選用手寫數字辨識、團體畢業照人臉提取和國畫人物提取等三種圖形識別問題進行實驗，並且比較辨識結果的準確率。實驗結果證明本論文提出的演算法能夠有效降低可變模型匹配的計算量，且辨識準確率優於硬式模板匹配。

Pattern recognition is an important issue in the field of computer vision. The difficulty of pattern recognition is how to integrate the image features and compute the image similarities so that the objective can be successfully identified. Template matching is a technique to compute the image similarities. Base on the difference of matching methods, template matching can be divided into two categories: rigid template matching and deformable model matching. Rigid template matching is faster, but its accuracy rate is lower than that of deformable model matching. We present a two-stage template matching algorithm which possesses the advantages of these two approaches. First, we use the rigid template matching to find out the candidate images which are likely to be similar to the objective, and then compute the similarities between candidate images and the objective using deformable model matching. After sorting the result, we achieve the purpose of pattern recognition. In this paper, we experiment on three different pattern recognition problems: handwritten digit recognition, face recognition of a graduation photo, and character recognition of a traditional Chinese painting, and we calculate the accuracy rate. The experimental result shows that the algorithm can successfully reduce the amount of computation task required in deformable model matching, and has a better accuracy rate compared to rigid template matching.