應用霍氏模型記憶幾何規則

Abstract

本論文為手寫字辨認設計兩項新的幾何規則，並將原有的特徵表示法做推廣，在應用霍氏模型的方法中得到較好的配對結果。除了手寫字辨認以外，我們也提出許多符合現今需求的新應用，本論文所提供的方法能夠繼續發展以解決這些更複雜的應用。 曲橢圓特徵模型適合用來表現複雜的幾何型態，這些特徵點在幾何上擁有許多規則，我們可以應用霍氏模型記憶這些規則以達成特徵點的配對。設計更多不同的幾何規則能提升識別率，也能解決許多不同類型的型態辨識問題。

There are two similarities constructed in earlier works, named inter-feature similarity and inter-link similarity. This work rewrites geomet-rical relations and constructs them into Hopfield model to improve the matching result. Constructing a set of bended-ellipse features is an efficient way for sampling cursive patterns. It converts a pattern into structural features and pro-vides a natural way for handprinted characters recognition. The idea which called feature-to-feature adhesion constructs a topological configuration for a pattern and transforms it into an undirected graph. The compatibility asso-ciated with the graph is then formulated as an optimization problem and is solved by a devised Hopfield network.