Research On Text Detection Algorithm In Natural Scene

The image text area in the natural scene contains a lot of useful information,which provides a certain foundation for scene text analysis.And text detection algorithms in natural scenes are also the basis for many computer vision applications.For example,obtaining license plate information,traffic sign information,etc.This paper will focus on the detection of text detection algorithms in the horizontal and arbitrary directions in the natural scene.The main tasks completed are as follows:1.Text detection algorithms in natural scenes.The first is the introduction of image preprocessing,including image preprocessing and target saliency of the image.Then it elaborates the feature extraction algorithm of text candidate regions,including the stroke width transformation and the most stable extreme value region.Finally,the relevant layers of the deep neural network and the forward and backward propagation algorithms of each layer are introduced in detail.2.Text detection in the horizontal direction in natural scenes.For the text region in the horizontal direction,this paper proposes a method based on the vertical region regression network.If we use the Region Proposcal Network method in the Faster-Rcnn framework algorithm to detect the text,there will be some restrictions.On the one hand,because the text area has variable length,complex background,diversification and other factors,the network must design a larger receptive field;on the other hand,the choice of positive samples during the training phase cannot be measured by traditional object detection methods and need to be trained.The stage selects the positive sample in the vertical direction based on the case that the ratio of the intersection between the ground-truth and the candidate frame(Anchor)in the vertical direction(Intersection-over-Union,IOU)is greater than a certain threshold.Regress the positive sample.Finally,multiple adjacent Anchors are combined into a text area.This method can make more small candidate boxes to cover the text area,thus improving the recall rate of text detection,and effectively solving the text area with variable length.The experimental results show that the recall rates on the horizontal datasets of ICDAR2011 and ICDAR2013 reach 0.815 and 0.826,respectively.3.Text detection in any direction in a natural scene.For text regions in any direction,based on Faster-rcnn,this paper proposes a training algorithm based on multi-angle text regions.There are two main improvements: first,improve the RPN framework to generate text candidate boxes of different orientations and scale sizes;second,in the second phase of Faster-rcnn training,use affine transformation algorithms to improve the original ROI(region of The mapping method of interest layer is proposed,and A-ROIPooling layer algorithm is proposed.The algorithm is beneficial to deal with various multi-directional text areas.The experimental results show that the algorithm has certain feasibility and validity.The recall rate and F1 score on the MSRA-TD500 dataset can reach 0.868 and 0.781 respectively.