Studies On The Detection And Recognition Of Scene Text

Text itself contains a wealth of semantic information,in addition to being a medium for people to communicate,but also widely used in nature scenes to help people understand their environment.Building signage is often used for navigation is a good example.In the computer field,text in images is also the key to the computer's understanding of image content.Such as detecting and recognizing watermarks in pictures,readings on industrial meters,street signs on roads,car license plates,price tags for different items in shopping malls,etc.The ability of text detection and recognition in these natural scenes is very helpful for realizing autonomous driving,building a smart city and improving people's work efficiency.Therefore,detecting and recognizing text in a natural scene has always been an important research problem.Different from the scanned document with the regular typeset black and white characters,the text in the natural scene has more manifestations,which has no standard plate,no fixed text font,no fixed language and no guaranteed picture quality.Besides,background may interfere the recognition of the text.Therefore,the detection and recognition of text in a natural scenario,despite its wide range of application scenarios,still present great difficulties and challenges.This article focuses on the detection and recognition of natural scene text.To explore a more robust neural network structure to enhance the precision and recall of text detection,and explore a more universal and robust method of text recognition.Specifically,this study includes the following aspects.1.Multi-scale Feature Extraction Based Text Detection Method: Due to the multi-scale of text itself,it is more sensitive to scale than general object detection.Therefore,this paper presents a natural scene text detection method which integrates multi-scale feature detection into the network design.By introducing the Inception structure,this feature layer of the network has more scale information,which makes the whole network more robust in the face of drastic changes of text scale.The method is tested on multiple datasets.Experiments show that this method may detect multi-scale natural scene texts,and may effectively improve the precision and recall rate.2.Feature Pyramid Based Text Detection Method: In view of the scale problem of texts,this paper proposes a text detection method which adds the network structure of feature pyramids to the neural network design so that different scale texts are trained and tested on different feature layers of neural networks.Experiments show that this method may detect multi-scale natural scene text,and may effectively improve the recall rate.3.Multi-scale Feature Extraction and Feature Pyramid Based Text Detection Method: In view of the scale problem of text,this paper proposes a method of text detection based on multi-scale feature extraction and feature pyramid.Which makes the network more robust in the face of the change of text scale.Experiments show that the method of recall rate reached the industry-leading level.4.Attention Based Text Recognition Method: In view of the problem that existing character recognition methods need fixed input image size,inspired by the language model,this paper proposes a method of applying the encoder-decoder structure with attention mechanism to character recognition.In the experiment,this method is combined with the ”Multi-scale Feature Extraction and Feature Pyramid Based Text Detection Method”and applied to the actual project,which achieves excellent recognition accuracy.