Research And Realization On Automatic Text Location And Recognition Algorithm Of Low-contrast Image

Character recognition has a wide range of application scenarios,such as anti-cheating,street view annotation,license plate location,and video caption recognition.To identify the text in the image,we need to locate the text area.Generally,the higher contrast of the image,the easier positioning of the text.But the low-contrast text image has low contrast,as well as its varied of text size,character type,and colors,which makes it difficult to locate the text area.This paper aims at researching the position of low-contrast images,proposed a algorithm based on AdaBoost classifier and regional similarity.Using this algorithm to achieve the automatic positioning low-contrast images with text size,Chinese and English,multicolor text.After that using the locating results to do some recognition work.This thesis firstly summarizes the domestic and foreign status of text localization and character recognition in images.Secondly,it describes the theoretical knowledge of text localization.Then,based on the difficulty of automatic localization of low-contrast image texts,it proposes a algorithm based on AdaBoost classifier and regional similarity: First,low-contrast images are subjected to pre-processing such as graying,contrast enhancement,edge detection,and edge information enhancement;then,training AdaBoost classifier with statistical characteristics based on edge and stroke width,using the connected region analysis method,the trained AdaBoost classifier,moving rectangle bar to gain the text regions,after that,using impoved binary method?region extension and combination method to deal with text regions.In order to verify the effectiveness of the proposed algorithm,we tested more than 1933 low-contrast images collected on the Internet to calculate the recall rate and miss rate of text localization.Compared with the existing methods,we verified that the algorithm can position the low-contrast images accurately and completely with different sizes of text,Chinese and English,multi-color.After completing the automatic positioning of the text area,this paper continues to identify the targeted text.First use the position coordinates to extract the text area from the image to get the text area image,and then use the improved binary method,erosion,region filling and other methods to preprocess the text area image,next use the improved vertical horizontal projection method to split the text into a single text character,finally,using neural network to train and recognize the single text character.