Handwritten and machine printed text discrimination using an edge co-occurrence matrix

Abstract

We employ an edge co-occurrence matrix (ECM) to distinguish handwritten and machine printed text without resorting to line or word information. The ECM is a modified co-occurrence matrix (CM) on edge images. First, the whole image is divided into overlapping range blocks with fixed size. Then, the ECMs are abstracted from these blocks. The ECM only counts the co-occurring edges connected with each other and its up direction part is the part with most distribution. The liner Support Vector Machine (SVM) is used to classify the features. Because of the similarities of neighboring blocks, the Discriminative Random Fields (DRF) is used to further improve the classification accuracy. The experiments on document images taken from HIT and IMA databases show the effectiveness of our proposed method.