Chinese Text Detection In Natural Scene Images Based On Deep Learning

With the development of information science and technology,various imaging devices have entered thousands of households,generating a huge number of digital images.The extraction of text information in digital images is a very challenging task and a problem that needs to be solved urgently.In recent years,the study of text detection in natural scene images has gradually developed in the direction of convolutional neural networks.However,the international research is mainly based on English environment images of natural scene.Compared with the English text,the Chinese text has many unique features in terms of texture,structure and distribution.The Chinese text detection of the natural scene should be carried out as an independent research field.In this thesis,Chinese texts detection are divided into horizontal direction texts detection and arbitrary multi-directional texts detection,and the Chinese text detection of different alignment directions is studied separately,the main research contents are as follows:(1)In this thesis,we analyzed the collection conditions of Chinese text and the natural scenes in which texts may exist,and fully understood the different characteristics of Chinese text and English text.We established a diversified natural scene Chinese text image dataset which contains 3600 text images in total,and labeled the Chinese texts information in each image which was used to train and test the Chinese text detection method in this thesis.(2)We proposed a horizontal Chinese text detection method based on YOLOv3 model.In this paper,the horizontal Chinese text is marked as a whole target.The K-means++ clustering algorithm is used to cluster the datasets.Then the clustering results are used to optimize the YOLOv3 detection model to achieve horizontal Chinese text detection in natural scenes.The experimental results show that the Chinese text detection algorithm has achieved good performance and real-time performance.(3)In this thesis,we adopted Res Net-101 network as the feature extraction network of EAST model to detect Chinese text in oblique direction.The EAST model can detect multi-directional text.In this thesis,we use Res Net-101 to replaced Res Net-50 as feature extraction network in the EAST Model,and test the Chinese text detection effect of EAST models under these two networks.The experimental results show that the EAST model based on Res Net-101 has achieved better results in multi-directional Chinese text detection.At the same time,it is found that the model is prone to over-segmentation in Chinese long text detection,which requires further research.