A Robust Text Segmentation Approach in Complex Background Based on Multiple Constraints

Natural Scene Text Understanding

“…Select a metric M to assign pixels to a cluster Choice of color space: colors to be clustered have different locations in different color spaces. Hence, k-means clustering in NS text extraction is either performed in RGB [66,80,147], in HSI [42,157], in YCbCr [39] or in a dynamically decorrelated color space using principal components analysis [24].…”

“…For k-means, few papers proposed solutions. Fu et al [39] included text geometry property after clustering to improve results such as in [144] with consideration on how to combine clusters to get sharp and consistent text components. Regarding GMM, spatiality is usually included by using the Potts model (also named Markov Random Field (MRF)) as GMM parameter resolution, as in [16,159].…”

Text Line Aggregation

“…A substantial amount of work has been carried out in text region localization in everyday scenes (Pilu, 2001;Chen and Yuille, 2004;León et al, 2005;Liu and Samarabandu, 2005;Liu and Samarabandu, 2006;Fu et al, 2005;Merino and Mirmehdi, 2007;Retornaz and Marcotegui, 2007;Lintern, 2008;Jung et al, 2009;Zini et al, 2009;Epshtein et al, 2010;Zhang et al, 2010;Pratheeba et al, 2010;Chen et al, 2011;Yi and Tian, 2011;Pan et al, 2011;Neumann and Matas, 2011a;Neumann and Matas, 2011b;Merino-Gracia et al, 2011), with many of them explicitly dealing with the text aggregation stage, such as (Pilu, 2001;Retornaz and Marcotegui, 2007;Epshtein et al, 2010;Neumann and Matas, 2011a;Chen et al, 2011;Pan et al, 2011;Merino-Gracia et al, 2011), although often other terminology was used for it, such as word or line formation. We now focus our review on these specific works, especially as we use several of them for comparative analysis.…”

A hybrid text segmentation approach

“…The latter extract text pixels by clustering pixels with homogeneous color into regions. For example, Fu et al [5] used the K-means clustering algorithm in YCbCr color space to generate several layers, and heuristic rules were employed to select text layer. Wang et al [6] clustered text images into some maps by color and scale of text strokes, and constructed a probability model online to identify text pixels.…”