Research And Realization Of Text Detection And Recognition In Natural Images

The scene text detection and recognition technology is a hotspot in the field of computer vision.A large number of researchers have carried on the long-term exploration on the direction.The result of the text extraction technology research in the natural scene can be applied to the intelligent urban traffic system,supporting system for the blind,unmanned technology and industrial automation systems.Different from the traditional optical character extraction technology,the natural scene contains a wealth of scene information.The scene text extraction is faced with the font variable,complex background,low image quality and many other difficulties.It has important practical application value to effectively improve the accuracy of the scene text extraction.This thesis discusses the research background and significance of scene text detection and recognition,describes and analyzes the present condition of scene text detection and recognition technology.The thesis focuses on the algorithm research of scene text detection and recognition.Scene text detection algorithm based on GPU and scene text recognition algorithm based on deep learning are designed.The image and video processing platform is independently developed and the scene text detection and recognition algorithm are verified.This thesis has mainly done the following work:In the aspect of scene text detection,this thesis discusses the existing traditional stroke width transform detection algorithm and analyzes the existing problems of the algorithm.A improved stroke width transform algorithm is proposed and the improved algorithm is designed based on GPU in order to improve the performance.The optimization strategy is used to optimize the performance of the algorithm and the speedup ratio of the final algorithm is more than 742.In the aspect of scene text recognition,this thesis studies the convolution neural network and applies Alex Net network to scene text recognition.This thesis improves the Alex Net network and extends the scene text data set by using the sample extension strategy.The feature extraction ability of Alex Net network is used to train a large number of data samples and the support vector machine is used to classify the extracted sample features.The recognition rate of Alex Net network is improved to 95.7%.This thesis designs and develops an image and video processing platform based on Open CV which includes the research results of scene text detection and recognition algorithm.The platform uses Direct Show plugin as video processing interface to realize the scene text detection and recognition in video.The proposed algorithm in this thesis is tested and verified in the film video subtitle extraction.Experiments prove that the proposed text detection and recognition algorithm are robust,the algorithm can process the natural scene video in real time.The proposed algorithm can automatically extract the text information in the video.At the same time,the proposed algorithm can also be applied in the extraction of the film subtitles.The algorithm can accurately locate the subtitle location in the movie video and recognize the subtitles.The proposed algorithm in this thesis has strong research and application value.