Research On Irregular Text Detection And Recognition In Natural Scenes

Scene text detection and recognition has become a very active research topic in recent several years.On the one hand,as it is fundamental in extracting textual information embedded in natural scenes,text detection and recognition methods in natural scenes have high research value.On the other hand,high-performance text detection and recognition systems are also of great practical significance in many application scenarios,including image search,real-time translation and robot navigation.Aiming at the shortcomings of the existing text detection and recognition methods.The research of this dissertation mainly has two aspects: one is to study the detection algorithm of multi-directional and irregular text in natural scenes.The second is how to make full use of the position and visual information of the text in the text recognition network to improve the accuracy of text recognition without context.Therefore,this dissertation studies from the following aspects:(1)A text detection method for natural scenes based on embedded re rating mechanism.In order to solve the problem of false positives in natural scene text detection,this dissertation proposes a natural scene text detection method based on embedded re-scoring mechanism,introduces the case segmentation network(Mask R-CNN)as the basic framework,designs the case mask re-scoring mechanism,and completes the detection of multi-directional and irregular text in natural scene.(2)Research on scene text recognition based on location information enhancement.In the attention mechanism based encoding decoder framework,some location or visual information is often ignored in the common long-term memory network(LSTM).This dissertation proposes a location information enhanced encoding decoder(PIE)framework for scene text recognition,which makes up for the deficiency of LSTM network by an additional location information enhancement module.(3)In view of the text detection and text recognition model proposed in this dissertation,we have carried out sufficient experiments on multiple comprehensive scene text datasets.Experiments show that compared with the mainstream methods in various fields,the proposed text detection and recognition method is more practical.