Text Detection In Nature Scene Images Based On MSER

Scene text detection refers to the detection of text areas in images captured in natural scenes.The text information in the scene image is an important semantic information,which is of great significance to the understanding,analysis and retrieval of the image content.Therefore,the scene text detection technology has important application value in the fields of intelligent transportation,industrial automation,multimedia retrieval and so on.However,due to the complex scene background,uneven lighting,and various fonts,the traditional optical character recognition technology can not be applied to scene images.This thesis employs the improved MSER algorithm combined with convolutional neural network for scene text detection.The improvement of the traditional MSER is mainly reflected in two aspects.On the one hand,in order to enhance the effect of fuzzy border image detection,gradient amplitude enhancement processing is used to enhance the text boundary.On the other hand,in order to improve the regional character proportion of candidate region,we apply a combination suppression strategy to filter out repeated regions,approximately repeated regions and nested regions.After using the improved MSER algorithm to obtain candidate regions,the Char-CNN classifier is designed to classify the candidate regions.We first extract the candidate regions of the characters,then use the hierarchical clustering algorithm to merge the candidate regions,and finally generate the text position information.We evaluate the algorithm performance on the ICDAR2013 data set in this thesis.Firstly,the improved MSER algorithm is tested.The results show that the improved MSER algorithm increases the recall rate of the character region from 88.9% to 90.2%,and the proportion of character regions in the candidate regions increases from 3.25% to35.19%.Then we evaluate the Char-CNN classifier.Results show that the classification accuracy of Char-CNN is 93.6%,which achieves 4.1% higher compared to SVM methods.Finally,the overall performance of the scene text detection algorithm proposed in this thesis is tested.The recall and accuracy rate of the algorithm are 0.68 and 0.85 respectively,and the F-Measure value is 0.76.Compared with existing scene text detection algorithms,the proposed algorithm has a relatively good overall performance.