Using LDA Model for Document Image Segmentation
Jilin Yang · Jisuanji fangzhen · 2011
Aiming at the problem of document image segmentation,we proposed a topic model based method to segment the document images into several areas,such as text,background,tables and figures.In the past,the segmentation of document images focused on threshold based method or supervised learning method.In our work,we first built a codebook using PCA and K-means which need only an unsupervised learning method.Then,the document images were coded using codebook and the probability of each code was calculated using LDA based method,which was followed by a Markov random field based labeling procedure.The advantages of proposed method include:unsupervised learning phase,more reliable probability from LDA,and smoothed segmentation results from MRF.