Research On Localizing Text Region In The Image

Text region localization is to locate text region in the image by using the methods of image processing. Because text in the image carries important information to describe and understand the image content, text region localization has become a hot research direction in the field of image analysis and processing in recent years. In this paper, localizing text region in the image under complex background has been studied specifically as follows:1、 For the problem of region in complex background which has similar structure to text will produce interference, a method based on region analysis and feature classification is proposed. At the stage of region analysis, edge detection is performed in channel RGB channels respectively to obtain an edge image, and connected component analysis is used to roughly determine text candidate regions. At the stage of feature classification,the histograms of oriented gradients of candidate regions are extracted as gradient features and local binary patterns of candidate region are extracted as texture features, and an adaptive and uniform local binary patterns is proposed. Experimental results show that the method can reduce the influence of similar structure in the complexity background and locate text region accurately.2、Aiming at the classic stroke width transform relies on results of edge detection and text is distinguished only in accordance with the region rules will lead to inaccurate positioning, a method that the stroke width is got by distance transform in maximally stable extremal region and text regions is classified by stroke features is proposed. Firstly, to overcome the impact of image blur, the image contrast is enhanced to generate bright text image and dark text image. And then the stroke width is confirmed by distance transform in the most stable extremal regions. Then non-text area is filtered out by stroke features extraction and classification. Finally adjacent text is connected together to form a line by text aggregation. Experimental results indicate that compared to the classical stroke width transform, modified stroke width transform can locate the text regions more accurately.3、In order to the accuracy of classification based on single feature image segmentation is not high, a kind of text classification based on graph cut model which is made of unary features and binary features of maximally stable extremal region is proposed. First of all, the most stable extremal region is detected as text candidate region. Then the gradient feature extraction, center-surround histogram and the coefficient of variation of stroke width is extracted as unary features, the color distribution and regional similarity is extracted as binary features. And then graph cut model is made of unary features and binary features, text classification is gained by the best optimal segmentation. At last, the final positioning is obtained by text aggregation. Experimental results show that the text location method based on graph cut model which is constituted by multi features can improve the accuracy of text classification and text localization.