An Efficient Approach for the Identification and Localization of Texts in Images

Objectives: To reduce false positive rate of text localization and to improve the performance of image segmentation process for better detection of text in images. Methods: Text information in images is an important consideration for image based applications such as automatic geo coding, content based image retrieval and understanding scene. But to detect the text from a complex background with different colours is a complex task. There are different techniques has been proposed to detect text in images. One of the techniques is hybrid approach for text localization and text detection in natural scene images. In this approach, the text region detector is used to detect text region and to segment the candidate components by local binarization. Then the non text components are filtered using Conditional Random Field model (CRF). Finally the text components are grouped together as line or words. This approach feels hard to segment some complex images due to lack of colour information. In order to enhance the performance of text detection in images Mahalanobis Distance (MD) metric, cosine based similarity metric and text recognition is introduced. Findings: The Mahalanobis Distance (MD) metric, cosine based similarity metric are computed for image segmentation process where colour information of images is considered instead of gray level image information. Based on the color information the input images are segmented. Then CRF is employed to filter out the non text components. After that text components are grouped together to localize the text from images. Moreover in the proposed work, text recognition is done by Kohonen neural network and compares the results of text localization and text recognition to reduce the false positive of text localization. Improvements: The experimental results show that the proposed work provides better results in terms of precision, recall, F1 measure and ROC curve.