Research And Implementation Of Scene Text Detection And Recognition Based On Target Detection Network

As one of the important forms of human-to-human communication or human-toobject interaction,text is an important element that can provide information.In recent years,detecting and recognizing text from scenes has become a hot research direction.Its purpose is to convert text images in scenes into character texts through algorithms,this conversion can be applied to many practical applications.Compared to traditional optical text recognition,the scenes text detection and recognition based on deep learning can adapt to more complex scenes without special requirements.This type of text detection and recognition technology can be applied in ticket recognition,content screening,and other applications.For example,it can provide convenience for ticket checking at stations and security supervision of information content for national security.Therefore,text detection and recognition technology has strong research value.In this thesis,through the study of related technologies,an end-to-end text detection and recognition network is designed.The text detection module is based on the target detection network and shares the features with the text recognition module to realize mutual supervision of the module training.The network can complete the task of text detection in scenes,it can detect text in any direction in complex scenes and translate text images into characters.The research content and results of this article are as follows:1.Improve YOLOv3 target detection algorithm.Corrected the problem of detecting long text lines,reduced the number of network layers,and accelerated the detection speed.At the same time,the residual network was used to provide the network with a shared feature layer to cope with complex scenarios.2.Design an end-to-end text detection and recognition network.The network generates the shared features of the detection module and the recognition module through the residual network.The detection module performs text area prediction based on the target detection network combined with the shared features,and the recognition module combines the text region features and shared features of the detection module to perform text recognition.Such a solution reduces the time and size of model training,and avoids the problem of lower recognition accuracy due to the difference between the two networks.This network can complete the task of text detection and recognition.Compared with non-end-to-end networks,it can extract more general features in feature extraction.The text detection and text recognition network modules can monitor and adjust each other to make the network parameters Better,better results.3.The text recognition module of the network is implemented by using an encoder that combines CNN and RNN and a decoder of CTC,and completes the task of translating text images with input sequences larger than the output sequence into characters.4.The operation of affine transformation between the text detection module and the text recognition module is used to unify the shape of the feature map input to the text recognition module,so that the feature map sent to the recognition module has a uniform height and is convenient.Perform character recognition.5.Implementing network training and testing based on the Darknet deep learning framework,testing the network model on different data sets and comparing experiments with different network models.The results of the experiment found that the network model designed in this thesis can cope with a variety of complex natural scenes,has strong robustness,and the algorithm can accurately detect the text in scenes and convert the text image to the text characters.The network have reached an excellent level in accuracy and recognition,and have strong research and application value.