Stroke Filter for Text Localization in Video Images

Two-stage approach to extracting visual objects from paper documents

“…The mean detection accuracy after first iteration was equal to 54 % (with highest 80 % for text and lowest 14 % for signatures). Observed low accuracy is caused by high resemblance between classes, e.g., many logos were classified as stamps, large number of tables (which according to [6] should be considered as graph- 15 Exemplary documents used in the experimental part ics) as printed text. The low accuracy for signatures comes from the lack of signatures in input documents; hence, we included the samples from SigComp2009, which are quite different in character.…”

“…Other authors made use of stroke filters [14][15][16], cosine transform [17] and LBP algorithm [18].…”

Fast and effective text detection

“…For instance, embedded text usually goes with dense edges [1]; text pixels almost have homogeneous color [2]; character strokes form distinct textures [3] [4]. Liu et al [5] analyzed the properties of character strokes and proposed to use stroke filter to directly detect and localize text. Based on different features, there are two main categories of modeling methods.…”

“…Based on different features, there are two main categories of modeling methods. One is constraint-based methods, such as [1][5] [6] [7] et al; the other is learning-based methods, such as neural network [8] [9], SVM [3][4] [5] [10]. The methods in the first category require designers themselves to finish the task of summarizing the effective classification rules, so it is not easy for engineers to create text detector with good performance.…”

“…In this paper, Inspired by [5], we utilize the stroke filter to obtain the low-level representation of image content, and then derive more compact classification feature to create a fast and effective text detector in a machine learning way. Compared with [4], our algorithm investigates to use more simple and distinctive features, so it has higher computation efficiency.…”