| Video surveillance has already covered every corner of the modern city, constructing acomplete well-structured surveillance system. The related surveillance technology plays avital role in preventing illegal and criminal activities, building the smart cities andimproving the quality of people’s life. The mainstream method resorts to computer visionand machine learning technology, which could cope with the long time surveillance frommultiple cameras and numerous video data.Text, an important expression of human language, appears everywhere in people’sdaily life. It is natural that the surveillance video contains lots of information about text,which could be used to detect or track the objects. However, the scenes which thesurveillance camera captures are usually natural scenes, text in which might be in variouscolors, sizes, fonts, orientations and other unpredictable factors, which brings greatdifficulties for detection and recognition. This paper focuses on the text detection in naturalscenes. The main work includes the following three aspects:First of all, this paper reviews the recently published methods, and makes acomparison between them.Secondly, the main framework of the text detection could be divided into two parts:component detection and text line construction. In component detection, this paper proposesa method which derives from several recent published methods. This method is connectedcomponent based, segments the image using Gaussian pyramid MSER (Maximally StableExtremal Regions). The candidate components are classified as text or non-text regions viarandom forest upon the features extracted by SFT (Stroke Feature Transform) and TCD(Text Covariance Descriptor). This method achieves a recall of0.75on ICDAR2013dataset regarding to component detection.At last, in text line construction, this paper proposes a novel method called Stable TextLine (STL), which constructs the graph with the detected components as the nodes, thedistance between the nodes as the weights of the edge, then applies multiple thresholds tocut this graph, picking the optimal text lines from several candidates based on the assumption that the components within the same text line resemble each other. Meanwhile,it is also natural to extend this method to detect the text lines in multiple orientations. Theexperiment demonstrates that the method proposed in this paper achieves agreeable resulton ICDAR2013dataset and could be compared favorably with the domestic leadingmethod. |
