Research On Arbitrary Shape Text Detection Based On Edge Contour Feature

As a carrier of information transmission,text exists in our daily life in a variety forms.With the continuous development of society,more and more application scenarios need to extract text information from images.In recent years,with the continuous development of deep learning technology,scene text detection has broken through the original development bottleneck,playing an important role in human-computer interaction,unmanned driving,intelligent detection and retrieval.Now it becomes one of the hot research directions in the field of computer vision and pattern recognition.However,there are still two serious challenges in the field of scene text detection.The first is the scale difference of scene text,and the second is the false positive problem in the process of arbitrary shape text detection.To solve these two problems,the present paper carried out research on arbitrary shape text detection based on edge contour features according to the characteristics of text images,and improved the two-stage method of scene text detection respectively to improve the detection ability of the model for text edge contour.The main research work and contributions of the present paper are as follows:(1)Aiming at the problem of scene text scale difference,a scene text detection model based on improved region proposal network was proposed.On the one hand,the point selection method of traditional RPN was improved to extract more abundant local region information of text edge,so as to improve the detection range of single text instance.On the other hand,Vertex-based VIOU loss function(Vertex-IOU)is proposed to improve the IOU value between the predicted boundary box and real boundary box better and faster by directly optimizing the normalized distance between the vertices of the predicted boundary box and real boundary box.Compared with the traditional smooth L1 loss function,it achieves better performance.Comparative experiments on multi-directional text data set ICDAR2015 show that the proposed model has better comprehensive detection ability compared with other models,which proves the effectiveness of the proposed model.(2)To solve the problem of false positives in the process of arbitrary shape text detection,the present paper integrated Canny algorithm into the traditional Ro I head and proposed an edge contour detection model based on deep learning.Compared with the original method,this model can detect more text edge contour.Moreover,NMS algorithm and double threshold algorithm can be used to further distinguish the text edge from the non-text edge,so as to suppress the false positives generated by the non-text edge.The experimental results on multi-direction text dataset ICDAR2015 and arbitrary shape text dataset CTW1500 show that the proposed model can improve the detection accuracy compared with other models,which proves the validity of the proposed model.