Text Detection Using Edge Gradient and Graph Spectrum Jing Zhang and Rangachar Kasturi Department of Computer Science and Engineering, University of South Florida {jzhang2, R1K}@cse.usf.edu Abstract In this paper, we propose a new unsupervised text detection approach which is based on Histogram of Oriented Gradient and Graph Spectrum. By investigating the properties of text edges, the proposed approach first extracts text edges from an image and localize candidate character blocks using Histogram of Oriented Gradients, then Graph Spectrum is utilized to capture global relationship among candidate blocks and cluster candidate blocks into groups to generate bounding boxes of text objects in the image. The proposed method is robust to the color and size of text. ICDAR 2003 text locating dataset and video frames were used to evaluate the performance of the proposed approach. Experimental results demonstrated the validity of our approach. 1. Introduction With the increasing availability of low cost portable photo cameras and video recorders, there are large and growing archives of multi-media data, such as photos and videos. These archives are useful only if they can be navigated efficiently. However, how to index and retrieve the information in multimedia data efficiently is still a challenging problem due to the semantic gap [1]. Fortunately, there is a considerable amount of text objects occurring in images and videos. As a well- defined model of concepts for humans’ communication, text embedded in multi-media data contains much semantic information related to the content. If this text information can be harnessed, it can be used to provide a much truer form of content–based access to the image and video data. Many supervised text detection approaches have been proposed in recent years due to the rapid development of classifier techniques. However, the performances of these approaches depend on the quality and quantity of training text samples significantly, which may limit their application on text with distortion or text from different languages. In this paper, we propose a new unsupervised text detection approach based on Histogram of Oriented Gradient and Graph Spectrum. The advantages of the proposed method are: (1) Text localization is based on the inherent properties of text edges, so it is robust to the size, color, and orientation of text; (2) Graph spectrum can group character blocks efficiently by capturing global relationships of text features. The rest of the paper is organized as follows: Section 2 reviews the related work. Section 3 describes and analyzes the proposed text detection approach. Experimental results are given in Section 4. Finally, we draw conclusions in Section 5. 2. Related Work According to the features used and the ways they work, text detection approaches can be divided into two categories: region based and texture based. Region based approach utilizes the different region properties between text and background to extract text objects. Color, edge, and connected component are often used in this approach. Shivakumara et al [2] adopt arithmetic mean filter, median filter, and an edge-based block growing method to extract text objects. The centroid of each edge is computed to measure the straightness which is used to eliminate background edges. Kim et al [3] use color and orientation consistencies to detect static text region in video. Liu et al [4] use an intensity histogram based filter and an inner distance based shape filter to extract text blocks and remove false positives whose intensity histograms are similar to those of their adjoining areas and the components coming from the same object. Bai et al [5] use a multi-scale Harris-corner based method to extract candidate text blocks. The position similarity and color similarity of Harris corners are used to 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.968 3963 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.968 3983 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.968 3979 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.968 3979 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.968 3979