Research On Multi-orientation Text Detection Algorithm In Natural Scene Based On MSER

The text in the natural scene contains high-level semantic information related to the surrounding environment.This semantic information has guiding significance in the areas of driverless,instant translation systems,and security monitoring.Text detection is a necessary step for image content analysis and the basis for text recognition.However,it is still a great challenge to extract text from natural scene images.On the one hand,the text in the scene image differs in size,arrangement direction,and language type.On the other hand,not only many backgrounds similar to the character structure exist in the scene image,but also factors such as blur,low contrast,occlusion,and lighting will affect the extracted text.Aiming at the above problems,this thesis conducts research from horizontal and arbitrary orientation text detection algorithms in natural scenes.The main research work is as follows:1.Aiming at horizontal text regions in natural scenes,the thesis proposes a text detection algorithm based on adaptive learning of Maximum Stable Extremal Regions(MSER).The algorithm is mainly divided into two stages.In the candidate region extraction stage,in order to solve the problem of low recall and low accuracy of the MSER algorithm caused by the blur and lighting.Firstly,the thesis use the edge enhanced MSER algorithm to extract the candidate regions of the text,and then use multi-mechanism suppression strategy to filter coarsely non-text.At the stage of non-text fine filtering,a candidate region verification algorithm based on Support Vector Machine(SVM)adaptive learning is proposed.Firstly,the training data set is extracted with three features of histogram of oriented gridients feature,uniform local binary patterns feature,and color perception difference feature.Secondly,the fusion strategy designed in thesis fuses three features,then the fused feature is used to train a support vector machine.Finally,the classifier trained filters out thoroughly the non-text of candidate regions.The experimental results show that the recall rate on the horizontal data set ICDAR2003 reaches 80%,and the accuracy rate is 81.7%.2.Aiming at the problems of poor text detection in complex scenes and difficult extraction of arbitrary orientation text,the thesis proposes a text detection algorithm based on the combination of MSER candidate region suggestion and convolutional neural networks.In order to solve the problem that the lower contrast and lower resolution of certain characters in the gray channel can easily lead to miss detection.Firstly,with the help of text color information,the thesis use the improved MSER algorithm to extract respectively richer text candidate regions in the three channels of R,G,B.Simutaneously,the candidate regions of the three channels are merged as the final text candidate region.Secondly,the method uses a two-class convolutional neural network to filter out non-text of the candidate regions,and then the algorithm generate a tilt positioning boxes with an angle based on the least squares linear regression algorithm.Finally,the thesis use the improved non-maximum suppression algorithm to filter out redundant positioning boxes.Experimental results show that the proposed algorithm can extract effectively text about any direction in complex images.The recall rate of the horizontal data set ICDAR2013 reaches 85.4%,the FPS is 15.4.The recall rate of the inclined data set ICDAR2015 reaches 81.3%,the accuracy rate is 83%,FPS is 8.25.