Research On Scene Text Of Arbitrary Shapes Detection Based On Deep Convolutional Neural Network

Automatic text detection based on scene image is the first step of optical character recog-nition system,and it is an important guarantee that optical character recognition system can be applied in all walks of life.Scene text detection technology has become a research hotspot in the field of computer vision and pattern recognition.It is widely used in daily life and work,such as scene understanding,product retrieval,automatic driving,map making,image and video surveillance and document retrieval system,etc.It provides great convenience for people's life and work.In addition,with the explosion of big data and the improvement of computer comput-ing ability,deep learning has developed rapidly,and has landed rapidly from academic research to industry.The advantage of deep learning is that it does not need to design the characteristics of a specific task by hand.It can automatically learn the features from large data according to the task.At the same time,the features acquired have better robustness and recognition perfor-mance.Because of the above advantages of deep learning,the current scene character recogni-tion methods based on deep learning have progressed substantially over the past years,and have achieved good performance in a controlled envirornment where text instances have regular shapes and aspect ratios.However,limited by the receptive field of convolutional neural network and the simple expression of text obj ects such as rectangular box and quadrangle adopted to describe text,previous methods may fall short when dealing with more challenging text instances,such as extremely long text and arbitrarily shaped textTo address these two problems,we present a novel text detector,which localizes the text progressively for multiple times.The proposed method consists of a direct regressor(DR),an iterative refinement module(IRM)and a shape expression module(SEM).At first,text propos-als in the form of quadrangle are generated by DR branch.Next,IRM progressively perceives the entire long text by iterative refinement based on the extracted feature blocks of preliminary proposals.Finally,a SEM is introduced to reconstruct more precise representation of irregular text by considering the geometry properties of text instance,including text region,text cen-ter line and border offsets.The state-of-the-art results on several public benchmarks including ICDAR2017-RCTW,SCUT-CTW1500,Total-Text,ICDAR2015 and ICDAR17-MLT confirm the striking robustness and effectiveness of the proposed method.In single scale testing,the pro-posed method achieves 62.3%in Hmean,surpassing the best single scale method RRD by 6.6%in long text benchmark ICDAR2017-RCTW.In curved text benchmark SCUT-CTW1500,the proposed method achieves 78.4%in Hmean,surpassing the best single scale method TextSnake by 2.8%.