Research On Methods Of Traffic Text Detection And Recognition Based On Deep Learning

Traffic text is a type of natural scene text.Unlike general scene text detection,traffic text not only has the characteristics of diversity and variability,complex background and occlusion,and poor imaging quality which natural scene text has,but also has its own characteristics,such as sparse distribution,slender text lines,and chaotic text layout.Traffic text detection and recognition is an indispensable part of current smart transportation and autonomous driving.Therefore,the research on traffic text detection and recognition is both full of great challenges and great prospects,I take traffic text as the research object and conducts research from two aspects: text detection and text recognition:1.Aiming at the characteristics of sparse traffic text distribution,slender text lines,and chaotic text layout,I designed a unique text area label based on EAST,which divides the reduced text area into a boundary area and a middle area.The boundary area is divided into a head pixel area and a tail pixel area.In the prediction result,the head pixel area is used to predict the coordinates of the two vertices on the left or top of the text box,and the tail pixel area is used to predict the coordinates of the two vertices on the right or bottom of the text box.This cleverly solves the problem of long text detection in traffic texts,and I test the performance of this method in public natural scene text detection through experiments,and also analyze the effectiveness of this method in traffic text detection through experiments.2.For the text feeding into the text recognition model is usually in an irregular shape,it may be that the poor performance of the previous text detection algorithm which causes the text detection box to not tightly surround the text,or the angle of view of the camera is skewed which causes the text to perspective distortion,or curve text that exists widely nowadays,these texts are presented in two-dimensional distribution on the picture,so this paper proposes to add a spatial transformation network(STN)in front of the text recognition network to correct the distorted text.We have determined the parameters of spatial transformation through experiments,and comparative experiments have also shown the difference in the performance of text recognition networks with and without correction network.3.In the training process of the text recognition network,in view of the huge number of text recognition classifications,we first pre-trained the proposed network on the abundant Synthetic Chinese String Dataset,and then based on the existing traffic text detection dataset,I made a traffic text recognition dataset.On the self-made traffic text recognition dataset,some distortion factors were deliberately added by random rotation,and then my model was fine-tuned on this dataset to further enhance the recognition ability of the text recognition algorithm in this paper on distorted traffic text.