Research On Text Detection Technologies In Natural Scene Images

Natural scene text detection,aiming to predict and locate the position of text in natural scene images,has been a hot research topic in the field of computer vision.Influenced by the factors of complicate background and text diversity,text detection still has weaknesses such as poor robustness and low precision,even some progress is achieved after more than a decade of development in the text detection field.In order to improve the performance of text detection,this paper do some research on two processes of text detection: candidate text regions generation and text regions location,which achieve better result.The main contents is shown as follows.(1)In order to solve the problems of low text detection performance and missed detection in existing candidate text regions generation methods based on Maximally Stable Extremal Region(MSER),firstly object proposal is applied to locate candidate text regions,and then combined with MSER to generate full initial candidate text regions.Finally,heuristic method is designed to filter regions which is detected by MSER and object proposal,to generate the final candidate text regions.Experiments show that the combination of object proposal and MSER can integrate the missed detected candidate text regions of MSER,to get better text detection result.(2)In order to solve the problems of poor performances of arbitrary text detection,a text detection algorithm based on random area grow algorithm is proposed in this paper.Firstly,stroke width,perceptual divergence and histogram of gradients at edges are designed to calculate the feature value,and a Bayesian algorithm is applied for their integration.Secondly,a conditional random field(CRF)model is used to label text region and non-text region based on the above feature values.And then,random area grow algorithm is designed to connect the character together which belong to the same word and text line.Finally,minimum-area encasing rectangle is used to locate the arbitrarily direction text,which make the final location results.Experiments show that,our algorithm achieves well performance on the classics natural scene text dataset,which is able to locate arbitrarily direction text.(3)According to the requirements of projects,a natural scene text extraction system is designed and implemented with MATLAB,based on the proposed candidate text regions generation algorithm which is combined of object proposal and MSER,and arbitrary directional text detection algorithm based on random area grow algorithm,and the open source text recognition algorithms in the world.The system is able to locate and recognize text contents from the input scene images,and have been accepted by the project with good results in applications.