Open AccessNatural Scene Text Detection Based On Multi-level Fusion Proposal Network
Author(s) -
Tong Li,
Wanggen Li,
Nannan Zhu,
Xuecheng Gong,
Jiajia Chen
Publication year - 2020
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/790/1/012051
Subject(s) - computer science , feature (linguistics) , artificial intelligence , bounding overwatch , minimum bounding box , text detection , recall rate , pattern recognition (psychology) , fusion , convolution (computer science) , feature extraction , semantic feature , natural (archaeology) , data mining , image (mathematics) , artificial neural network , philosophy , linguistics , archaeology , history
Natural scene text detection is a challenging issues. In this work, a multi-level features fusion two-stage text detection network was developed to solve the problem of the insufficient use for feature map and the difference between text and common target. In order to obtain deep semantic feature, the network framework of VGG-16 was improved, and different levels of convolution feature map were confused, meanwhile the fusion approach was partially adjusted. Also, vertical proposal network was used to classify and regress bounding box. Results indicated the accuracy and recall rate of this method are 85.4% and 81.0%, respectively, by evaluating the net on ICDAR2013 dataset. The experimental results suggest that the multi-level features fusion method can improve the efficiency of feature map in natural scene text detection.