Research And Implementation Of Natural Scene Text Detection Based On DenseNet

With the rapid development of social level and the improvement of computer processing level,it has now entered the era of data explosion.Benefiting from this,computer vision has derived many branches,including various fields such as face detection,image retrieval,security monitoring,smart cars,scene text detection and recognition.This subject is based on the "Internet Content Supervision Platform-GA" project of Kaitian Communication Company,the 30 th Research Institute of China Electronics Technology Group,and researches on text detection and recognition of natural scenes.In this thesis,the traditional Optical Character Recognition(OCR)technology is directly applied to natural scene text detection and the results shows that the robustness is very poor.However,the existing natural scene text recognition method based on deep learning has many problems in practical application and its accuracy rate and recall rate indicators still need to be improved.This thesis focuses on the scene text detection,and then gives a detection method for natural scene text based on an improved Densely Connected Convolutional Networks(Dense Net)in the tilt direction for the detection problems.The main content of scene text detection in this thesis includes the following three aspects: First,the basic network Dense Net is first improved to extract the features of natural scene text.The improved Dense Net model structure can extract deeper features to solve the problem of insufficient feature extraction of traditional neural networks.Second,a Dense Layers layer is newly designed for the frame regression and text prediction.In the mean time,for the frame regression,a densely connected multi-scale prediction module is designed to accurately detect the position of the scene text.According to the characteristics of the scene text,the quadrilateral frame is used for regression to enable it to detect the scene text in the oblique direction.Third,different from the traditional post-processing method,this thesis uses Soft Non-maximum Suppression(Soft-NMS)for post-processing.Compared with the traditional Non-maximum Suppression algorithm(NMS),the post-processing method of the Soft-NMS algorithm is better,without adding additional parameters and training amount,which is easy to implement.Based on the design idea of natural scene text detection in this thesis,the research content of scene text recognition is as follows: First,the improved basic network(Dense Net)is used to extract features and it can extract more detailed features.Second,connect a cyclic layer behind the convolution layer to obtain deeper context information and a sequence of feature vectors.Third,a transcription layer is connected behind the circulation layer to identify the label distribution of each frame.Finally,this thesis tests the above methods.In this thesis,the text detection and recognition methods for natural scenes based on Dense Net in the tilt direction are compared with the existing algorithms on the horizontal dataset and the tilt dataset respectively.The test results show that the detection method proposed in this paper obtains better results.