Research Of Scene Text Detection Based On Deep Learning And Sample Expansion

The spread of information is extensive and rapid,and images are increasingly appearing in information interaction.The text information in the image has the research value that can't be ignored.By detecting and extracting the text information in the image,the image can be classified according to its content.With the development of artificial intelligence technology,scene text detection has gradually become a hot spot in the field of machine vision due to its complex background and more realistic significance.In this thesis,the scene text detection algorithm based on deep learning is used.Traditional scene text detection algorithms,such as MSER and SWT,are easily affected by uneven illumination,image noise,complex background and et al.And they are troubled by error accumulation when generating and eliminating character candidates.The scene text detection algorithm based on deep learning combines these above steps into one step,which can effectively avoid the problem of error accumulation.The main work of this thesis includes:firstly,most of the scene text detection algorithms are trained on tens of thousands of images,which requires a large number of scene images.For the scene images without annotation,manual annotation is required,which requires a large amount of work in the data preparation process.In this thesis,a sample expansion method based on scene image is proposed to simulate the scene text pictures under different spatial angles,spatial positions and other conditions,and generate multiple synthetic scene images.The sample expansion algorithm uses 229 training set images of ICDAR2013 data set as the basis to synthesize new scene text images through color space change,text region transformation and background region transformation.Then,object detection SSD method was analyzed.SSD method used feature map of different levels to make joint prediction,integrated detection process into a single step,and improved detection accuracy.Based on SSD object detection method,this thesis proposes a text detection model based on SSD.SSD text detection model achieves good results in ICDAR2013 dataset.Finally,based on the scene image sample expansion method and SSD text detection method,a SSD text detection method based on sample expansion is proposed.In combination with the idea of weak supervised learning,only part of the synthetic scene image is marked completely.And this part of complete annotated data is combined with the original small sample data set to form the initial training set.The positive data samples suitable for training in the coarse-grained annotated and unannotated composite images,which are selected by SSD text detector to the training set for supplementary training to improve the text detection effect.The experimental results show that the text detection effect can be improved through sample expansion in the small sample training set.Compared with the text detection algorithm using the large training set,the recall rate and F score are also improved to some extent.