Research On Text Detection In Complex Natural Scene

With various application and rapid development of smart mobile terminals and the Internet,the natural scenes can be recorded and shared in form of pictures,which contain abundant information,the most important one is the text information in natural scenes.That information serves in the fields of language translation,Navigation,multimedia information retrieval and many other specific applications.Text detection and recognition are both the key steps in the above applications,the detective accuracy is often linked with the efficiency and universality of the image application.Because of the complex background,the changing characters,easy-influenced by light or shooting orientation,there are many uncertainties in the text detection under natural environment.Therefore,it will be of great significance to study technology of texts detection in natural scenes.This paper focuses on the algorithm of text detection in natural scenes.The algorithm is improved from the following three aspects: the extraction of candidate connected regions,the formation of candidate text regions and the classification of text non-text regions.(1)Applying the maximum stable extreme region to extract candidate connected regions,in order to reduce time and computational complexity of subsequent work,correctness of text classification,this paper proposes a new method to smooth the maximum nesting;(2)Stroke width reduction,combined with the maximum stability extremum,is used to select the candidate regions,and the lost character recovery algorithm is proposed to detect multi-directional texts;(3)Advantages and disadvantages of AdaBoost and SVM are fully considered in text/ non-text regions classification.The AdaBoostSVM concatenation classification algorithm is proposed.Combined with depth learning,text detection is carried out by convolutional neural network.In order to detect text detection more effectively in complex scenes,it is necessary to set scale and scale in text candidate regions.And the convolution integration layer can obtain more features in small regions.This paper,improves the text detection algorithm in several aspects.In addition,the proposed method detects the horizontal direction on ICDAR2013,a dataset commonly used in text detection,and conduces multi-directional detection on MSAR-TD500.The experimental results show that the proposed algorithm can detect any direction,and at the same time,target texts also achieved good results in images with complicated background and uneven illumination,which significantly improved the accuracy and recall rate of text detection.