Research On Scene Text Detection And Recognition Based On Boundary Fitting

Scene text detection and recognition is one of the research hotspots in the field of computer vision,which has a wide range of application prospects in intelligent driving,automatic production,text image retrieval and other scenes.For text detection in natural scenes,the existing mainstream detection methods are mainly based on segmentation.However,this method relies too much on the accuracy of the boundary obtained from the classification map,which may cause incomplete text region detection and more background noise.In scene text recognition task,due to the complexity and variability of text,the existing methods face the poor robustness of text recognition network.Therefore,this thesis conducts an in-depth study on the problems existing in the above two tasks,improves the detection method based on segmentation and the recognition method based on attention mechanism,studies how to make the text detection box better fit the text shape,how to obtain more accurate recognition results,and builds a two-stage scene text detection and recognition model.The main work of this thesis is as follows:(1)A scene text detection method BMINet based on boundary fitting is proposed to improve the accuracy of text positioning.In order to separate the compact text,direction and distance information are introduced as priori knowledge.At the same time,this thesis makes an in-depth analysis of the problem that the detection area cannot contain complete text content or contains too much background noise,and designs two effective modules.One is the boundary fitting module,which makes the detection box fit the text shape better through the control point offset.The second is multi-scale fusion module,which fuses multi-scale feature map information to increase network receptive field.Experiments show that the proposed module can effectively improve the performance of scene text detection network and perform well in multiple data sets.(2)In order to improve the accuracy of scene text recognition,this thesis proposes a double-supervised scene text recognition method with text region correction.Correction network is introduced for text region correction,and two decoding branches are used to supervise the network.This method integrates the recognition methods based on CTC loss and attention mechanism.It not only pays attention to the textural information of text images,but also learns the context dependence relationship,and improves the accuracy and robustness of network recognition by means of dual supervision.(3)From the perspective of verifying the performance of branch detection and branch recognition,this thesis designs an image processing module that integrates the rotating rectangle box and the segmentation binary graph,connects the improved scene text detection method and recognition method into a two-stage model through the image processing module,and improves the correction network.The experimental results show that the two-stage model can achieve good detection and recognition effect,and the recognition accuracy can reach 80.6% on the Total-Text data set.