Research On Key Technologies Of Scene Text Localization And Recognition Based On Deep Learnin

Scene text detection and recognition aims to detect and recognize text from complex natural scenes.Unlike simple scenes,scene text detection and recognition is affected by illumination,blur,and occlusion,and it still faces great challenges.In view of this,this paper investigates the deep learning-based method for detecting and recognizing text in scenes,and the research includes the following two points:(1)For the autonomous unmanned vehicles,this paper proposes an efficient end-to-end scene text detection and recognition model based on a two-stage object detection model.In order to make full use of the features extracted from the backbone network,this paper proposes a novel feature fusion method,which can fuse text features and features of the text surrounding area to further improve the performance of the model.This paper proposes a novel two-branch structure to address the issue that the object detection model cannot predict both text and character.In addition,due to the lack of character-level annotation in public datasets,this paper uses a training method based on weakly supervised learning.This method can effectively transfer the character prediction ability learned by the model on synthetic datasets to real datasets.(2)This paper incorporates a visual representation learning method based on Momentum Contrast(Mo Co)to enable the primitive representation learning-based scene text recognition model(Primitive REpresentation learning Network,PREN)to learn better character features,thus improving the performance of the model.In order to achieve contrastive learning on character-level,this paper creates character-level feature queues,classifies and stores character features.Furthermore,in order to further improve the robustness of the learned character features,this paper uses computer vision library to render a noise-free regular text image and uses this image to expand the positive sample type of contrastive learning.