Research On Text Detection Of Traffic Signs In Complex Natural Scenes

Nowadays,the natural scene text detection has became a popular researching direction,and traffic sign text is one kind of natural scene text,it plays an important role in the visual perception of the driverless system.Although promising results have been achieved in the areas of traffic symbols detection,little attention has paid to text detection on traffic signs.In fact,the text on traffic signs contains more abundant and valuable traffic information than the traffic symbols,if it can be combined with traffic symbols,it will bring more benefits to the development of related applications.In this work,we firstly propose an effective scene text detector for Chinese traffic signs,whose pipeline only consists of a Fully Convolutional Network(FCN)and a simple post-processing step.A specially designed pixel location method of head and tail text makes the detector have robust detection results for those words with long scale and sparse arrangement.The proposed method achieves the F-score 0.79 on Chinese traffic sign text dataset(CTST-1600)which is built by our research group,and also achieves competitive results on the ICDAR 2013 and MSRA-TD500 when compared with other state-of-the-art methods.The experimental results show that our method has great adaptability to both traffic sign text detection and other complex scene text detection.In the experiment,we find that the method in Chapter 4 performs weak when dealing with the complex light conditions and its network speed isn't fast enough.In view of the above problems,Chapter 5 puts forward the corresponding improvement scheme,which is embodied in: an image preprocessing module is added to head of the original network,which can detect the intensity of light of the input image and automatically enhance the image judged as poor light,so as to effectively improve the recall rate of the text detection model in the complex scene conditions;A scale transfer layer is designed to replace the unpooling layer in FCN,this technology can expand the feature map and compress the channel to 1 / 4of the original channel number at the same time,so as to reduce the amount of network parameters and calculation and speed up the network.The improved text detection model improves the F-score of detection on CTST-1600 from 0.79 to 0.82,and FPS from 4.90 to 5.28.Compared with the other three methods,the improved method shows obvious comprehensive performance advantages.