Research On Text Detection Algorithm In Natural Scene

The text in life contains rich semantic information,so the text detection in the natural scene has a broad application scene and important practical significance.However,text detection in natural environment will be affected by many factors such as complex background,lighting,font,tilt deformation and so on,so the detection effect is not very ideal.In recent years,with the development of deep learning theory,many excellent detection algorithms have been proposed to solve this problem.EAST is one of the most excellent algorithms.Nevertheless,there are still some problems in EAST,such as insufficient sensitivity field and unreasonable sample weight allocation,and there is room for further improvement in accuracy and speed.This thesis conducts indepth research on the above issues,and the main contents and achievements include:(1)a new training sample generation strategy is constructed.When the sample was generated,the clipping boundary was reduced from 0.3 times to 0.1 times,the text area was expanded,more edge information of text area was added,and the weight of the sample was adjusted accordingly.The random cropping strategy is improved,the screening conditions are relaxed,and the fault tolerance rate of positive samples is improved.Experimental results show that,without changing the structure of EAST network,the sample generation strategy proposed in this paper can improve the detection performance of the algorithm.Compared with the classical EAST algorithm,the recall rate,accuracy rate and comprehensive performance of the text detection open data set ICDAR 2015 have been improved to some extent.(2)Build a text detection algorithm based on improved EAST.On the basis of the classical EAST network,the ASPP(Atrous Spatial Pyramid Pooling)network was added to fuse the features of the receptive field of different scales,so as to increase the receptive field,reduce the complexity of the network and improve the training speed.Replace the class balanced cross-entropy loss in the classic EAST network model with dice loss,adjust the sample weight according to the text area,improve the attention to the text in the small area,improve the unreasonable sample weight,and speed up the convergence of the algorithm.The experimental results show that compared with classic EAST detection algorithm,The natural environment text detection algorithm constructed in this paper has significantly improved detection performance on ICDAR 2015 and ICDAR 2013 data sets while maintaining 15 FPS.Compared with several classical deep learning detection algorithms(CTPN,Seg Link,Pixel Link,etc.)proposed in recent years,the algorithm in this paper also achieves excellent comprehensive performance.