Research On Text Detection And Location In Complex Background Images

Text location in complex background images has always been a very difficult andchallenging research topic because of the text type variety and the complicated background.This dissertation mainly studies the Chinese and English text location in complex backgroundimages and proposed a solution which can locate the Chinese and English text from thecomplex background images effectively on the basis of summarizing the existing main textlocation methods.The text location method proposed in this dissertation is mainly divided into three parts:image segmentation, candidate text region generation, and text classification. In the part ofimage segmentation, we explore some existing algorithms for image segmentation andpropose an improved image segmentation algorithm based on mean shift. The experimentalresults indicate that the proposed method for image segmentation could segment the text andbackground effectively. Meanwhile, the background image is homogeneous after imagesegmentation which makes the amount of connected region less, thus the complexity of thetext extraction process is reduced.In the phase of text regions identification during the candidate text regions generating,three kinds of constraint are proposed: the constraint based on the text corner information, theconstraint based on axial projection, and the constraint based on that the text would not appearon the image boundary. In addition, a brush algorithm is proposed in the text region mergingphase. The proposed algorithm could merge some scattered characters into a complete Englishword or incorporate different parts of the Chinese character into a complete Chinese character.Experimental results show that the proposed method for generating candidate text regionscould obtain the candidate text regions quickly and completely.Aim at the different characteristics of English and Chinese, we respectively extracted thedifferent kinds of features for text classification while using the AdaBoost classifier in theprocess of text classification. In the process of English text classification, we suggest twodifferent strategies for extracting the histogram of oriented gradient (HOG) feature and thelocal binary pattern feature (LBP), and compared their classification performance. Then HOGfeature and LBP feature are combined together as the English text classification features.Moreover, we extracted the images’ Gabor feature and six kinds of texture features thatinclude mean, standard deviation, energy, entropy, inertia, and local homogeneity feature forChinese text classification, and compared the test results with HOG and LBP features’classification results. Experimental results show that the extracted features can characterize the text of English and Chinese, so that the text location method proposed can locate the textregions in complex background images.We tested the proposed method for English text location by using two hundred andfifty-one images in ICDAR2003database, and the accuracy of text location is0.71, the recallof text location is0.65. In addition, we test the proposed method for Chinese text locationusing a self create Chinese text image database containing one hundred images, the accuracyof text location is0.72, the recall of text location is0.68. The results verify the validity of thismethod.