Design Of Text Font And Color Recognition System Based On Deep Learning

With the rapid development of digital technology,the rapid maturity of the Internet,digital media,video media,and other technologies makes visual elements occupy a large part of the information carrier,and fonts,as an important form of expression,are constantly developing and presenting diversification.form.At the same time,our country's protection of font copyrights is becoming stricter.In recent years,there have been many claims for illegal use of commercial fonts,with the number of claims ranging from hundreds of thousands to hundreds of millions.Therefore,the identification of fonts in pictures is very important and has great commercial value.Not only that,with the development of deep learning technology,document analysis technology has evolved from text recognition to more advanced document understanding,and digging out more text attribute features such as fonts and colors are of great help to layout recovery and document understanding.Based on a comprehensive study of the current situation at home and abroad,based on the existing deep learning technology,this thesis designs the font recognition algorithm module and the text color extraction algorithm module,combined with the more mature OCR(Optical Character Recognition).Technology and Flask background web framework finally built a set of text attribute(font,color)recognition system.The main research results obtained in this thesis are as follows:1)Propose a font recognition algorithm module based on metric learning.Collected 700 fonts,designed an automatic data synthesis algorithm to build a data set,built a metric learning network model structure to realize the mapping from the input image to the font features,and finally realized the classification of fonts through the metrics between features.While introducing the recognition framework,the accuracy of Top1 was further improved,reaching 69.282)Propose a text color extraction algorithm module based on a text segmentation network.The traditional color extraction algorithm combines image segmentation technology to segment the area to which the text belongs,which effectively suppresses the interference of the text background to color extraction.The network structure designed in this thesis has two branches,which are respectively the information supervision of the original image and the information supervision of the text area.The experimental results show that the network of this thesis only needs a small number of training samples to have good generalization performance.3)A text font and color recognition system was built.Introduce the mature OCR text recognition framework,realize the end-to-end system design from the picture input to the font and color recognition result output of each paragraph of the text in the picture,and finally complete the visual demo display on the Web site.