Research On Scene Text Recognition Algorithm Based On Deep Learning

With the development of artificial intelligence technology,it has been widely used in the field of computer vision.As an important subject in the field of computer vision,scene text recognition has attracted much attention in recent years.Scene text recognition is to detect and identify the text images under different natural scenes.It has a very broad research value and application prospect in the fields of industrial automation,automatic driving,image retrieval,and so on.However,most text images in actual natural scene usually have various deformations,complex font styles and different character scales,which largely increases the difficulty of scene text recognition.Aiming at some key points in scene text recognition,this thesis proposes two identification algorithms,which have greatly improved the recognition accuracy of natural scene text.The main works are summerized as follows:1.We propose a scene text recognition algorithm based on position information enhancement.Considering the problem that the existing networks rely too much on context information,this thesis proposes a location information enhancement branch to solve this problem.The algorithm mainly enhances the text correlation in the feature sequence,which is learned by convolutional neural network(CNN),through a correlation attention module,and then adds a location information to each feature sequence using a location enhancement module.Finally,the proposed location information enhancement branch is spliced with the coding sequence in the traditional network through a fusion decoder.Thus our network can learn enough features for text recognition while ensuring the amount of parameters.The experimental results demonstrate the effectiveness of each proposed module and the effectively improved identification ability of the overall model.2.We propose a scene text recognition algorithm based on hierarchical awareness and global enhancement.In which,the hierarchical network can learn the transformation of character scale well,and the global enhancement network can better learn the correlation between the characteristics of different network layers.The proposed network can effectively solve the problem of inconsistent feature information between CNN and transformer structures.A large number of experimental results show that compared to the commonly used connection modules,the recognition effect is greatly improved after adding the proposed network structure.