In this paper, we propose a text detection method that improves the feature map of Hangeul text area using MGAP (Modified Global Average Pooling) for accurate detection of Hangeul text area. TheCNN(Convolutional Neural Network) based text region detection algorithm has result of detect a background region and including a nontext element. In order to solve these problems, we proposed an MGAP method that applied multiple filters to improve the area where text features were detected in non-text areas.The feature map of text generated by MGAP also contains some background areas, so accurate text detection is not possible.To solve this, we separated the text and background into K-means clustering by using the feature that the color of the text is the same. The final text region was extracted by comparing the clustering result image and the feature map using MGAP.In the case of Hangeul, the problem that consonants and vowels are detected separately was improved by using the structural features of Hangeul. The experimental data set was experimented with KAIST data set consisting of Hangeul and English text images. The proposed method shows an improvement on 1.7% for the English data set, 2.5% for the Hangeul data set, and 1.8% for the English and Hangeul data set when compared to the existing text detection method.
Volume 11 | 07-Special Issue
Pages: 1826-1833