Research On Horizontal Text Detection Technology Based On Multi-feature Fusion

Detecting text from images refers to locating text area in image.It is widely used in image retrieval,intelligent education.The task originates from the requirement of mobile advertisements analysis.In this scenario,the text is almost horizontally distributed.It is necessary to quickly and accurately locate the text from complex background.The model should also meet the requirement of size limitation in the deployment,.In recent years,with the rise of deep learning,many excellent text detection algorithms have emerged.Compared with common objects,text has a sequential nature and is directional,corresponding network structures need to be design.In the application,the scale of text regions often changes greatly,the mainstream two-stage detection algorithms lack consideration of different scale detection scenarios.Algorithm need be designed to effectively extract and fuse sequence features and hierarchical features without affecting the detection time too much,reasonable multi-feature fusion method is also necessary.In the fragmentbased text detection algorithm,the detection results generated by the general detection method contain redundant area,a reasonable text line construction algorithm needs to be designed to generate more accurate detection result.The thesis focuses on the above analysis and makes full use of the features of text sequence.We have designs a reasonable and effective net work structure for text sequence features and multi-scale problems,the core of which is the fusion of features within different levels.The fusion method includes two dimensions: multi-scale feature fusion between layers and sequence feature fusion within layers.The purpose of the former is to solve text object detection in different scales,while the latter is to fully exploit the sequence characteristics of the text.Experiments show that fusion in these dimensions can improve detection results.In view of the shortcomings of the existing models in the post-processing stage,we propose an improved text line construction algorithm based on the maximum stable extreme region.The traditional method has a high false detection rate but can generate relatively accurate detection results.In the post-processing stage,this thesis uses the accuracy characteristics of the traditional method to design an improved algorithm that can remove redundant detection fragments.Aiming at the shortcomings of the model in the oblique text detection scene,the improved text line construction algorithm can also obtain oblique quadrilateral detection results which are closer to the real text area.In order to verify the feasibility of the ideas presented in this thesis,experiments are performed on the ICDAR2013 dataset.We have compared the algorithm with mainstream detection algorithms.The F score is 0.82,and detection time is about 0.12 seconds per image,which prove the validity of the ideas.Then we made tests on the self-built advertisement dataset,and the F-score is 0.84.Experiment results show that the algorithm proposed in this thesis can meet the requirements of practical applications.