Learning-based scientific chart recognition

In this paper, a learning-based paradigm for scientific chart recognition is proposed. Two kinds of chart recognition methods are presented: hidden Markov model based and neural network based method. A newly developed feature extraction method is also put forward for chart images. Experiments on three kinds of charts show that the ergodic hidden Markov models achieve a satisfactory result for chart recognition. Unlike traditional primitive-based diagram recognition method, learningbased approach need not recognize the graphic primitives in charts. Thus the method bypasses the recognition error problem caused by inaccurate primitive extraction that is also a major obstacle to the construction of a general chart recognition system.