Research On Text Detection Algorithms In Natural Scenes

Character is a tool of human communication.It is an expression of simple visual pattern to oral language.With the development of deep learning technology,Natural scene text detection has been one of the most important subjects in computer vision and model recognition.This paper makes research about text detection under natural scene.Different from the text pictures recognized by the traditional OCR optical character recognition technology,text under the natural scene pictures has distinguished features of inconsistent character size,uncertain text direction,complex background and so on,which make the detection more difficult and bring some unnecessary problems to the subsequent identification processes.In terms of the detection of text of natural scene,scholars around the world give excellent solutions.While,these solutions have poor accuracy because of the simple calculation or have complex parameter computation and these solutions have poor generalization of different data sets,so it is hard to balance robustness and accuracy.According to the problems above,this paper is aimed to understand and research the text detection task under natural scene based on the deep learning algorithm and traditional algorithm to research.This paper mainly does the following work:1.Data preprocessing and Transfer learning.The natural scene text image is the source of the rise of scene text detection technology.Because the existing public data sets have different characteristics,this article uses the image processing direction algorithm for the used data set in order to make the detection algorithm have better effect.For image normalization and image enhancement.In order to make the subsequent detection model converge faster,this paper intercepts the text area delimited by the real label in the MSRA-TD500 dataset as positive samples,and then randomly cuts the background area as negative samples to train Resnet-34 model classifies the background and text region.The backbone network of the pretrained model is used as the feature extraction network for subsequent detection models.2.Scene text detection algorithm combining traditional detection algorithms and deep learning methods.Through the comparative analysis of the text areas in the scene text image data set,this paper finds that most of the text written by people in order to express a certain idea is the same color.Therefore,based on the research,based on the Maximum Stable Extreme Value Region(MSER)algorithm,a regional gray level fusion algorithm based on the distance metric is constructed.This algorithm is used to remove small background regions and combine separate character regions to obtain candidate regions.In order to distinguish whether the candidate area is a text area,and to reframe the candidate area containing multiple text lines,and to balance the efficiency and accuracy of the text detection algorithm,this article improves the YOLO(You Only Look Once)target detection algorithm.Using the candidate area as separate image data to train the improved YOLO algorithm to remove the background area and relocate the text box to an area containing multiple text lines.3.Compare with different classic text detection algorithms on multiple public data sets.Through algorithm construction,experiments and research,experiments are performed on the ICDAR2013,ICDAR2015,and Ali Tian chi ICPR2018 datasets,and compared with the experimental results of multiple classic text detection algorithms,it shows that the natural scene text detection algorithm in this paper works on certain types of dataset has good robustness and good generalization to different datasets.This also provides a different idea for the subsequent research of scene text detection algorithms.The research and development of subsequent scene text detection algorithm has a certain reference value.