| Using machine to locate and identify in natural scene images by machine is not only can improve the efficiency of information processing,but also greatly simplify people's work,which can save a lot of manpower,therefore texts location and recognition technology in natural scene images has been used in various fields of production and life.Nowadays,texts localization and recognition technology in document is more and more mature,however texts localization and recognition in Natural scene is still a bit insufficient.Therefore,the location and recognition algorithm of text in natural scene is still a hot research topic.Based on the analysis of English texts and Chinese Texts of natural scenes,some new methods are proposed for the shortcomings of the three aspects,including the natural scene English texts location,the Chinese texts location and the Chinese character recognition in natural scene,therefore the main contents of this thesis are as follows:1.In view of the problem that existing algorithm of English texts localization in natural scene does not work well for the skewed texts and running speed of algorithm is slow,a fast algorithm of skewed English texts localization in natural scene based on Maximally Stable Extremal Regions(MSER)with hierarchical clustering is proposed.Firstly,the suspected text regions can be located in the images quickly by MSER elliptical fitting algorithm,then the non-text regions are quickly filtered out according to the self and spatial features of the fitting ellipses.By using the idea of hierarchical clustering,the missing non-text regions can be deleted while the text regions fuse from small to large layers,which can locate skewed English texts quickly and accurately in natural scene.2.In view of the problem that the current algorithm of Chinese texts localization is not effective for Chinese character stroke fusion,and the effect of filtering of non-text regions is not good according to the character of Chinese connected areas,an algorithm of Chinese texts localization based on MSER and Support Vector Machine(SVM)is proposed.Firstly,this algorithm locates the candidate stroke regions in the image quickly through the MSER pruning algorithm,then uses a dynamic closed operation algorithm based on Stroke Width Transform(SWT)to fuse the strokes of characters in the images effectively which can solve the problem of poor fusion of strokes.Secondly,extracting the 388 dimensional features of Gabor and Histogram of Oriented Gradient(HOG)from images of text regions as samples,and using the SVM after sample training to filter non-text regions,the expected results have also been achieved in the final test in self built Chinese location data set.3.In view of the problem that there is no connection between each pixel in the traditional character complexity Chinese characters recognition,an improved algorithm of Chinese characters recognition based on word complexity is proposed.Fristly,the text images need to be pretreated with 3 steps,including denoising,skew correction and skeleton extraction,secondly,obtaining the direction statistical histogram of text images by improved text complexity extraction method,using 8 dimensional feature datas in the histogram as the input of BP neural network,using fuzzy output as output of BP neural network.The BP neural network was trained with the training samples,finally,in the recognition test,the recognition rate of Chinese characters is greatly improved compared with the traditional Chinese character complexity recognition method. |
