Research On Text Line Detection Algorithm Of Rectangle Traffic Panels

Driverless systems and advanced driver assistance systems are important parts of intelligent transportation.Automatic detection and identification of road traffic signs is one of the key issues in driverless systems and advanced driver assistance systems.The rectangle traffic panels have plenty of road traffic information,such as texts,symbols and so on,which play an important role in improving traffic safety and traffic efficiency.Therefore,it has great research value and application prospect to quickly detect and extract the text information in rectangle traffic panels.Chinese texts within rectangle traffic panels are chosen as research object.Chinese texts within rectangle traffic panels are detected and text lines are extracted.The main work of this paper includes the following aspects:(1)The coarse detection of text regions in rectangle traffic panels.First,the method of combining RGB space and HSV space is adopted to preprocess the image in order to reduce the impact of color distortion and illumination change on text detection.Then,a multi-layered detection algorithm based on octree color quantization is given to achieve the coarse detection of texts in rectangle traffic panels.(2)The precise detection of text regions in rectangle traffic panels.Based on the coarse detection of Chinese texts,a multi-cascade filtering algorithm based on machine learning is designed to obtain the precise location of texts.Specifically adopting geometric features and HOG features combined with Adaboost classifier forms detection and filtering system to remove non-text regions.Finally,recalling text algorithm is designed and the lost texts are recalled to further improve the recall rate,then the Non-Maximum Suppression algorithm is used to improve the accuracy of text detection.(3)The detection of text lines in rectangle traffic panels.This paper improves metric learning based on kernel logistic regression to detect text lines for the precise detection of text regions.This method has good robustness to varying sizes,sporadic distributions,and different distances of Chinese text regions in rectangle traffic panels.In order to test the performance of the proposed algorithm,1509 Chinese rectangle traffic panel database is established,randomly selecting 436 images as test images,and the remaining 1073 training images.The detection recall rate of Chinese texts is 88.63%and the detection precision of Chinese texts is 95.81%.The detection recall rate of Chinese text lines is 89.94%and the detection precision of Chinese text lines is 93.26%.Experimental results show the effectiveness of our proposed algorithm.