Research On Text Detection And Multi-script Identification In Natural Images Based On Machine Learning

Text is an important carrier of human emotion and cultural heritage,it plays very important role in many aspects of human production and life.Text is a common element in the modern urban environment,such as posters,signs,billboards,etc.,which contain a large number of text.The text in natural scene can convey rich and accurate high-level semantic information,and it is the key element to understand the content of the scene.The script identification of text in natural images is an important direction of content-based image retrieval and multi-lingual system development.In a natural image,the detection and identification of text is very challenging:on the one hand,the text in a natural scene possessed rich diversity,it may various in color,number,size and space,and may belong to different languages;on the other hand,the background of natural scene is complex and there are some problems such as noise,occlusion and perspective,various factors have brought huge difficulty to the detection and identification of multi-lingual text.In now days,how to effectively deal with the natural images containing several kinds of scripts is an urgent problem to be solved.To solve this problem,text localization algorithm based on visual saliency and edge density and script identification algorithm based on basic image features and machine learning were proposed in this dissertation.Firstly,a text detection algorithm was proposed,combining the visual saliency and edge density.In proposed algorithm,multi scale spectral residual method was used to detect visual saliency regions,and Sobel gradient operator was employed to detect image edge in the saliency regions and then the edge density was obtained.After the preprocessing of image edge by morphological method,text areas were detected by means of prior hypotheses for text arrangement.Secondly,a multi-script identification method based on basic image features and machine learning was put forward.This method made the character sample set of Arabic numerals,English,Russian,Japanese Kana,simplified Chinese and Korean,and furthermore,the skeleton of letters was extracted.According to the structural characteristics of different languages,combined with the features of classifier,the language identification was divided into two stages:coarse classification and fine classification stage.In the coarse classification stage,text was divided into two categories using support vector machine.The first category includes Arabia numerals,English,Russian and Japanese kana.The second category includes Chinese and Korean.In the fine classification stage,support vector machine was used to identify the first class,and BP neural network was used to identify the second class.The experimental results show that the text detection algorithm achieves 73%precision and the multi-script identification algorithm achieves 73.33%precision,which proves the method proposed in this dissertation is effective and feasible.