Research On Text Detection Method Of Natural Scene Based On Deep Learning

Text in natural scenes can express advanced image semantic content.It is a key element of human understanding of natural scenes.Therefore,the research of text detection and text recognition of natural scenes is of great significance.Natural scene text detection and recognition technology has a wide range of application prospects in many fields such as unmanned industrial automation,intelligent security,and human-machine interconnect and many other fields.Text detection of natural scene is a very important pre-processing link in the optical character recognition(OCR)of natural scenes.All text recognition images are from the areas of text detection.In traditional optical character recognition,the processing object is usually an image with a relatively high resolution,and has a relatively simple background,a single color,and a regular layout,so the difficulty of text detection and character recognition is relatively small.Compared with traditional text recognition,text detection tasks in natural scenes are more complex and more challenging: On the one hand,the texts in natural scenes have diversity.For example,the size,color,font,direction and arrangement may all be different;on the other hand,such as the illumination intensity,resolution,noise and object obstructions and many other uncertain factors,which brings great difficulties to the text detection in the natural scene.This paper first introduces the natural scene detection algorithm based on the Maximally Stable Extremal Regions(MSER).which is not robust when dealing with complex background images.In this paper,a universal target detection algorithm(Single Shot Multibox Detector,SSD)is introduced to detect texts in natural scenes based on deep learning,SSDs cannot detect extreme aspect ratio text in natural scenes.The third and fourth chapters of this article have improved on this defect of SSD.In this paper,an algorithm of horizontal direction text detection in natural scene(Text-HD)and an algorithm of arbitrary direction(multi-reverse)text detection in natural scene are proposed.Both of which use a single network model to directly regress and classify targets.Text-HD designed the special aspect ratio for the text aspect ratio in the default boxes,and specially designed the "long" type(1*5)convolution kernel in the Textbox layer,which can be effective for different The horizontal direction text in the aspect ratio natural scene is detected.Text-OD changes the "long" convolution of the Textbox layer to 3*5.At the same time,it uses the method of regressing quadrilateral to replace the horizontal bounding box of horizontal text detection.It perfectly matches multi-directional text in natural scenes..In addition,it can achieve faster detection speed while ensuring a higher detection rate.The algorithm has no post-processing operations other than non-maximum suppression(NMS)processing.