Mathematical variable detection in scientific document images.

Mathematical expression detection in scientific documents is a prerequisite step for developing a mathematical retrieval system that has attracted many researches recently. In the detecting process, one challenging issue is the detection of variables. The similar properties of variables and narrative text cause many errors in the detection in existing approaches. In the paper, a novel detection methodology of variables in inline mathematical expressions is proposed. The merit of the method is that it can operate directly on the variable images without the employment of character recognition. The proposed method uses the features of projection profile of images and the fine-tuning of different machine learning algorithms in the detection process. The achieved accuracy varies from 86.14% to 94% for the detection of variables in inline expressions in document images in various public benchmark datasets. The performance comparison with existing methods demonstrates the effectiveness of the proposed method.