基於少量訓練資料之新穎形態相似度量於圖形識別

Abstract

電腦視覺在近十幾年來快速發展，圖形識別是當中一個重要的議題，而通常又被視為分類問題。根據處理方式不同，可分為基於記憶基礎與基於函數模型進行圖形識別，然而圖形識別問題的困難處在於如何克服組內的高變異度，過去的研究往往依賴大量資料來達成辨識目的。本論文主要探討如何於少量的訓練資料基礎上處理圖形辨識問題，並深入研究如何計算形態之間的相似度，提出一個兩階段的研究方法。我們首先於樣板資料庫建立階段，在特徵空間中選取類別中具有代表性的訓練資料，藉以降低記憶空間與運算時間負荷，第二個階段則利用形態匹配與模型轉換，藉以提升同類變異容忍度，最後結合不同的形態相似度量方法下，利用加權式K-最鄰近分類器達到分類的目的。我們選用手寫數字資料庫與MPEG-7圖形資料庫進行實驗，實驗結果顯示本論文提出的形態相似度測量方法能夠有效提升辨識準確率，達成圖形識別的目標。

Pattern recognition, often considered as a classification problem, is one of the most important issue in computer vision which has dramatically developed in the recent years. According to the different processes, it could be divided into two major categories: memory-based method and model-based method. Moreover, the major concern of pattern recognition is how to handle the high within-class variance of data. However, in the previous literatures, most studies would rely on a great amount of data to deal with it. The aim of this thesis is to implement a two-stage technique of pattern recognition based on the shape similarity measure with small training data. In the first stage, we would establish a prototype database by selecting representative training data in the feature space, which can reduce memory usages and consuming time. Furthermore, in the second stage, we would apply shape matching and modeling transformation to enhance the tolerance of within-class variance. Finally, a multiple shape similarity measurement is utilized to classification with distance-weighted KNN. In this study, MNIST handwritten digit database and MPEG-7 shape database are used for testing the technique, and the experiment results indicate that it could achieve the purpose of pattern recognition with higher accuracy rate.