Research On Scene Text Detection And Recognition Based On Neural Network

With the development of the Internet,there are more and more ways for people to obtain and share pictures from natural scenes,and it becomes more and more important to recognize text from massive natural scenes.The detection and recognition technology of text in natural scenes is also widely used in actual production and life,such as unmanned driving,automatic translation,intelligent assistance for the blind,and other fields,which have become a hot research topic in the field of computer vision.However,the text images in the scene have various shapes,complex backgrounds,and the influence of light at the same time,which increases the difficulty of recognition.Traditional detection and recognition algorithms cannot meet actual needs.A large number of studies have shown that neural networks can achieve better results in image feature extraction,which has become the current mainstream research direction.In response to the above problems,this paper proposes an arbitrary shape text detection and recognition algorithm based on neural networks.The main research content includes:(1)Aiming at the problem of inaccurate positioning of irregular text,based on the PSENet algorithm,an improved ISTDNet is proposed.This method includes a low-cost segmentation head and a post-processing algorithm for multi-scale expansion.The erosion and expansion algorithms in morphology are introduced.Irrelevant background is removed by the erosion operation,and the real text area in the foreground is expanded by the expansion operation.The ISTDNet text detection algorithm can more accurately distinguish the foreground and background of the text image.It is suitable for small text scenes and improves the detection accuracy of irregular text.The text location accuracy of the algorithm has been increased from84.2% to 85.6%.For closely adjacent text regions,a combination of multi-scale expansion algorithm and deformable convolution is used to improve the feature extraction ability,which can better distinguish adjacent text and improve positioning accuracy;(2)For irregularities For the problem of low accuracy of text recognition,the ITRNet text recognition algorithm is proposed.The algorithm is composed of two parts: correction network and recognition network.The text is corrected by a progressive multi-target correction method.On this basis,the corrected image is recognized through the recognition network based on CRNN+Attention sequence,and the text recognition is accurate The rate is 2.1% higher than the existing method.Experimental results show that compared with the performance of existing detection and recognition algorithms on the CTW1500 data set and Total-Text data set,the algorithm proposed in this paper can further solve the problem of irregular text recognition and detection,and the recognition effect has been significantly improved.