| Optical character recognition is a key technology and an important research direction in the field of computer vision.This technology is widely used in all walks of life,such as online marking,ticket recognition,license plate recognition,automatic driving,image retrieval,etc.The text in the image is in a variety of forms.The text size,tilt angle,bending degree,color change,text density,etc.will bring difficulties to optical character recognition.At the same time,image background,image shooting angle,light factors,etc.will also affect the recognition.How to accurately identify the character content in the image is still a problem to be solved.Optical character recognition mainly includes text detection and text recognition.Text detection is to find the text area in the image,and text recognition is to identify the content of characters from the text area.The thesis will carry out algorithm research from two areas of text detection and text recognition.The main work of the thesis includes:1.In order to enhance the recognition ability of the model for connected handwritten Chinese characters,a handwritten Chinese character dataset is created,and a text line dataset is generated from the handwritten Chinese character dataset.Handwritten Chinese character data set is composed of a single character sample set written by multiple writers.This data set covers all first-level Chinese characters,common punctuation marks,and uppercase and lowercase English letters.It can be used for training and testing the recognition of single handwritten Chinese characters by models.In order to facilitate the training and testing of the text line recognition task using the data set,the thesis uses the collected handwritten Chinese character data set to generate a text line data set,and provides scripting tools that can generate custom text lines,the script tool can generate any number of training samples and test samples based on user needs.2.To better detect the text region in the image,the thesis proposes WFF-DBNet algorithm based on the text detection algorithm DBNet.The algorithm adds an expansion convolution module to the feature pyramid module,and then a bidirectional weighted feature fusion module is designed to replace the FPN structure,and the spatial channel attention module is added.The experiment indicates that the suggested model can detect text regions effectively and has achieved outstanding detection outcomes on various public datasets.3.For the text recognition task,the thesis proposes the TRAR algorithm based on the classic text recognition algorithm CRNN.First,in order to improve the recognition ability of the model for curved text,the STN correction module is added to the model for text correction,and then the two-branch feature extraction module is introduced,and residual connection are embedded in the circular network layer.The experimental results on multiple data sets show that the algorithm has strong text recognition ability.4.Using the text detection model and text recognition model proposed in this thesis,a feasible text detection and recognition system is created.The system has the ability of text detection,text line recognition,end-toend text detection and recognition.At the same time,the system supports text detection and recognition in multiple scenarios.Based on deep learning technology,this thesis proposes corresponding algorithm models for text detection and text recognition tasks.The effectiveness of the algorithm is proved by many experiments,and it is verified and analyzed on the handwritten Chinese character data set created in this thesis.At the same time,it is also deployed in the actual application system.Therefore,the research of this thesis has certain theoretical significance and application value. |
