Text Detection And Recognition Of Network Image Based On Deep Learning

With the development of deep learning technology,the improvement of computing power and the growth of visual image data,visual intelligent computer technology is used in many application fields such as smart photo albums,face recognition,image search,urban smart traffic management,smart medical care,and smart cities.Both have achieved impressive results.In all these studies and applications,pictures are an important medium for transmitting information.There are a large number of important texts everywhere in pictures.These image texts concisely,clearly and clearly express the information exchanged between different groups of people.Therefore,with the rapid growth of online pictures,picture text recognition is very important and worth studying.Deep learning and neural networks have greatly changed and reshaped computer vision.This is another major leap in the field of science and technology following the revolution of steam technology and power technology in the history of human civilization.As an important research field of computer vision,network image text detection and recognition is inevitably affected by this wave of revolution,and then it enters the era of deep learning.In recent years,academia has made great progress in concepts,methods,and performance.The text in web images has become an active field.The reading of web image text is usually divided into two sub-problems: text detection and text recognition.This article aims to summarize and analyze the detection and recognition of network image text based on deep learning,mainly focusing on:(1)It is mainly to adjust the loss function of the YOLOv5 s object detection framework,and use the COCO data set to verify the accuracy and precision of the modified YOLOv5 s framework object detection;secondly,use the two data sets VOC2007 and VOC2012 to compare Fast R-CNN,Faster R-CNN,SSD,YOLOv3,YOLOv4,YOLOv5 s and other detection frameworks to conduct comparative experiments to further verify the detection performance of YOLOv5s;finally use the ICDAR 2017 and ICPR MTWI 2018 challenge data sets to detect text categories in network images,And give the corresponding bounding box and score to the detected text at the same time.(2)Mainly to improve the CRNN text recognition framework,modify the activation function of the first two layers of the CRNN network,use the CRe LU function to save the amount of calculation;use instance normalization(IN)and batch normalization(BN)The combination of replaces the original batch normalization(BN)layer.Two data sets of ICDAR2003 and ICPR MTWI 2018 are used to verify the text recognition performance of the improved CRNN framework.(3)Combine the object detection framework and the text recognition framework to finally form the text detection and recognition framework of the network image.The text in the image is detected through the object detection framework,and then the corresponding text area is extracted by the text extractor,and the text is passed into the text recognition after processing The module performs text recognition and finally outputs to the text label.The ICDAR 2017 and ICPR MTWI 2018 challenge data sets are used for model training and optimization,and the data sets composed of regular pictures and irregular pictures are used to verify the performance of the proposed framework to prove the validity and reliability of the framework.