Text Detection with Deep Neural Network System Based on Overlapped Labels and a Hierarchical Segmentation of Feature Maps
Autor: | Hong-Hyun Kim, Dong-Joong Kang, Jea-Ho Jo, Zhu Teng |
---|---|
Rok vydání: | 2019 |
Předmět: |
0209 industrial biotechnology
Computer science business.industry Value (computer science) Pattern recognition 02 engineering and technology Convolutional neural network Computer Science Applications Hierarchical clustering 020901 industrial engineering & automation Character (mathematics) Control and Systems Engineering Feature (computer vision) Salient Segmentation Artificial intelligence business Word (computer architecture) |
Zdroj: | International Journal of Control, Automation and Systems. 17:1599-1610 |
ISSN: | 2005-4092 1598-6446 |
Popis: | This paper proposes a three-level framework to detect texts in a single image. First, a salient feature map of text is extracted using a Fully Convolutional Network (FCN) that achieves good performance in semantic segmentation. Label combination using both boxes of word and characters level is proposed to improve the detection of uneven boundaries of text regions. Second, in the feature map of FCN, the text region has a higher probability value than the background region, and the coordinates in the character area are very close to each other. We segment the text area and the background area by using the characteristics of text feature map with Hierarchical Cluster Analysis (HCA). Finally, we applied a Convolutional Neural Networks (CNN) to classify the candidate text area into text and non-text. In this paper, we used CNN which can classify 4 classes in total by separating the background area and three text classes (one character, two characters, three characters or more). The text detection framework proposed in this paper have shown good performance with ICDAR 2015, and high performance especially in Recall criterion, finding more texts than other algorithms. |
Databáze: | OpenAIRE |
Externí odkaz: |