| Based on deep learning theory,this paper researches the method of text line detection in natural scenes.The current academic research in the field of computer vision has benefited from the rapid progress of deep learning technology,and has achieved promising performance in sub-problems such as object detection,semantic segmentation,and object recognition.Text line detection in natural scenes is a fundamental and important issue in the field of computer vision,and is a key part of many applications.The research method of this paper belongs to instance segmentation,and the main research contents can be divided into the following three parts:1)This paper summarizes the latest development of text line detection algorithm in current academic circles,introduces the defects and some improvement measures in convolutional neural network,which lays the foundation for the development of new methods.In view of the problem of text line detection,many improvements have been published based on the general target detection algorithm in the current academic circles.In this paper,these algorithms are classified and summarized.2)A feature enhanced network structure is proposed,CFPM.Starting from the basic FPN network structure,this research designs a stacked multi-level fusion structure to enhance the features.In addition,in order to better train the text line detection model,we use the synthetic data set to make the pre-training model,so as to achieve a good balance between the accuracy and rapidity of the text line detection task.Feature enhancement can also reduce the label data needed for training.The CFPM feature enhancement structure proposed in this paper can ensure high accuracy and near real-time speed without adding many parameters.3)A text line detection algorithm CFPM-DB+ based on AC loss is proposed.In this study,through analyzing the shortcomings of the loss design in the original DB semantic segmentation module,we use AC loss to improve the DB module.The model based on semantic segmentation can adapt to the input image marked by different shape boundary boxes,and can get the same size prediction graph as the original one.This method can reduce the tedious post-processing steps and accelerate the reasoning speed of the model.The CFPM-DB and CFPM-DB+ algorithms proposed in this study have achieved relatively balanced and comprehensive performance indicators on two open datasets. |
