Research On Irregular Text Recognition Algorithm Based On Deep Learning

With the continuous development of science and technology,the application of artificial intelligence such as deep learning is becoming more and more extensive,and computer image processing is an important part of it.As a key stage in computer image processing,the text recognition technology of natural scenes has been continuously developing in recent years.By detecting and recognizing text in natural scenes,obtaining the information people need in natural scenes has great value and prospects in many fields such as license plate recognition and human-computer interaction.The difference between text recognition in natural scenes and it in documents is that the background of natural scenes is more complex and diverse.At the same time,irregular text,such as uncertain arrangement directions and curved shapes which can be seen everywhere in natural scenes,makes it even more difficult.Focusing on the characteristics of text in natural scenes,this paper studied irregular text recognition algorithms in natural scenes,which are based on deep learning methods.The main content can be divided into the following three parts:1.Studied one scene text data synthesis algorithm.Scene text images get many problems,such as inadequate dataset,nonuniform distrubution of image quality,and inaccurate manual annotation,which have a major impact on the training of the text detection and recognition algorithm.In order to solve these problems,we have studied a scene text image synthesis algorithm for synthesizing the training data we need.2.For irregularly arranged texts in natural scenes,we made the text detection frame more suitable for obliquely arranged text,by introducing the angle parameters into the original horizontal rectangular enclosing frame on the basis of the one-stage detection framework.In addition,we integrated the anchor learning mechanism of the two-stage into our detection algorithm,by replacing the original anchor with the learned anchor into the final prediction,and now the improved algorithm can detect Multi-angle slanted text,achieving better text detection result.3.For irregularly shaped texts in natural scenes,an irregular text recognition algorithm basing on focus enhancement was studied.The convolution kernel of the traditional convolutional layer is generally a 3 * 3 rectangle,but irregular text may show a curved shape in the image.We used a deformable convolutional layer to transform the convolution network,making our text recognition algorithm concentrates on the text area to improve the ability of recognizing irregularly shaped text.In addition,our algorithm can recognize irregularly shaped texts without correcting.