Research On Text Detection And Recognition Algorithm In Natural Scene

In recent years,scene text detection and recognition has become a hot topic in the field of computer vision.Due to the popularity of mobile computer devices and smart phones,the acquisition of scene text image data is more convenient and efficient.Due to the existence of various forms of text,different resolutions,background noise and other factors in natural scene images,the traditional text detection and recognition methods are difficult to achieve good results,and the text detection and recognition methods based on deep learning have high research value and significance.The main research work of this paper is as follows:(1)On the basis of the PSENET text detection framework based on Res Net50,a semanticaware text bounding box is proposed,which imposes constraints on the pixels of the boundary of the text area.A branch is added on the basis of PSENET to predict the bounding box of the text area,thus effectively solving the problems of adjacent text instance adhesion and incomplete text detection in the PSENET text detection framework,and greatly improving the detection accuracy of the model.(2)Based on the PSENET text detection framework,a text area attention mechanism is proposed,which applies more attention to the text area than the non-text area,that is,greater weight,making the model more sensitive to the text area.The text area attention mechanism guides the detection model to further distinguish between text areas and non-text areas,thereby alleviating the problem of missed detection in the PSENET text detection framework and improving the recall rate of the model.(3)A thin plate spline transform network is proposed on the CRNN text recognition framework.The irregular text image is rectified into the normal horizontal text image through the thin plate spline transformation network,which improves the recognition accuracy of the model.Finally,through the comparative experiments on the scene text detection data set and the scene text recognition data set,the text detection and recognition method proposed in this paper is verified to be more robust and applicable.