Research On Irregular Scene Text Recognition Methods

In recent years,with the rapid spread of mobile devices and wide coverage of wireless networks,pictures and videos have replaced text as the main way people share their lives,resulting in a huge amount of image data.In natural scene images,text carries rich semantic information,providing an important reference for scene content understanding.There are a large number of irregular texts in the scene images,which is a difficult problem that many current models cannot effectively handle with.Irregular scene text recognition refers to text that suffers from shooting angle and spatial layout,such as arbitrary-oriented text,curved text and perspective text.This thesis analyzes the existing irregular text recognition methods and explores the problem of irregular text recognition from two distinct perspectives,coming up with corresponding solutions.This thesis proposes a novel text image rectification network for robust recognition of irregular scene text.Existing works mostly directly perform a global transformation on the text image,asking for a highly expressive deformation function,such as a thin plate spline transformation,which may make characters distorted.Based on the idea of space division and conquer,this thesis segments the text image and uses multiple simple local rectification functions to approximate the complex target deformation function.This thesis first predicts the upper and lower control points of the text line,and then uses two Bezier curves to fit the upper and lower boundaries of the text line respectively.Then denser control points are produced by uniformly sampling on the Bezier curves,and two adjacent pairs of them enclose a quadrilateral area.Next,a projection transformation function is applied to these quadrilateral text blocks and stretches them into rectangular blocks of fixed size.These rectangular blocks are spliced together in the original order to get the rectified text image fed to the recognition network.Next,this thesis proposes a novel spatial attention based irregular scene text recognition model,which differs from existing spatial attention methods in that,our attention model explicitly infers and exploits parameterized shape cues of the text for alignment of character features to simulate a person's reading process,who sequentially estimates the position of each character,recognizes it,and then moves forward the sight point the next character.Specifically,given an input image,this thesis infers the geometry of the text's centerline and further predicts the alignment position of each character on the centerline.A Gaussian kernel is then employed to compute the attention map of the current character to extract its local features used in a seq2 seq recognition module.This thesis conducts experiments on several public scene text recognition datasets to evaluate the effectiveness of the proposed methods.And the results of these experiments demonstrate the effectiveness of the two irregular scene text recognition methods proposed in this thesis compared with existing methods.