Research On The Location And Segmentation Of Unconstrained Text In Images

The recognition of text in static image and video(frames[i.e.static images]stream),is constituted by two stages:firstly,detecting and extracting the text in image,its aim at segmenting text section from input image,this stage is named as text detection;secondly,recognizing the detected text section,its aim at identifying separate characters as the corresponding text result from input text section,this stage is named as text recognition.In that,text detection and positioning which mainly used to determine the location of the characters in the image,and to find out the bounding box of the text,is the most critical step.The text segmentation shall remove as lots as possible the wanted background,for the sake of the subsequent text recognition.In computer vision,in order to achieve image processing,image analysis and image understanding,the text detection and positioning is essential as the basic step and critical phase.This is the significance of the study in this thesis.Literature researches show that the text recognition of natural scene images is difficult to directly apply the text recognition algorithms in the traditional standard(constrained)image,due to the different font sizes among characters and words in the natural scene image,the different directions,different typefaces,different degree of blur,different illumination,different degree of obstruction of obstacles;in addition,real-time requirements are relatively high for natural scene image.Any text/character is composed of strokes,and corner detection is the key to detect stroke.Among commonly used corner detection algorithms such as SURF,AGAST,BRISK,FAST,SIFT,ORB,the FAST(Features from Accelerated Segment Test)algorithm is not scale invariance,but with a certain degree of rotation invariance and affine invariance,more important is the speed was faster,this algorithm is more suitable for real-time applications.Therefore,based on FAST algorithm and stroke width conversion algorithm,an improved FAST detection algorithm(iFAST-improved FAST)-a fast text corner detection algorithm,is proposed to locate and segment the region with unconstrained text in the image.The iFAST detection algorithm first detects the corners of the stroke in image,extracts the text fragments from the corner attribute,and then uses the multi-scale adaptive pyramid model to train the cascade classifier to remove the extra non-text area.The algorithm can quickly,robustly and accurately detect and segment different size text areas in image.An effective text clustering algorithm based on text orientation voting is also used to collect the detected areas into the text line to allow subsequent stages(e.g.OCR modules)to be processed.Using the ICDAR and MSRA-TD500 data sets commonly used in the field of text recognition as training set and test set,and comparing with other algorithms,it is found that the proposed iFAST can achieve better performance in multi-directional text and multi-directional text.Compared with the commonly used MSER text detection algorithm,the iFAST detection algorithm reduces the number of text sections about 1/2 times,and can detect more than 25%of the characters,while the detection speed is 4 times higher.The iFAST detection algorithm with the subsequent classification stage can reduce the number of text regions to about 1/7 and 3 times faster than the MSER detection algorithm.