Hierarchical Text Detection In Natural Scene Images

Scene text detection is a method that detecting text regions from natural scene images. It is a challenging work because the texts in the natural scene images often appear with the complex background and a variety of text types. Furthermore, the quality of the natural scene images is very different. However, text detection from natural scene images has recently gained attention due to its potential application in various areas. In this work, we propose a novel text detection approach which is on the basis of previous work.Taking in consideration of the problems existed in the previous works, the proposed novel text detection method includes three aspects:(1)A character region extraction algorithm based on multi-scales is proposed.Some bigger characters are difficult to be detected in natural scene images.To solve this problem,we proposed a character region extraction method based on multi-scales.The comparative experiment shows that the proposed method can detect the bigger character effectively.(2)A character region extraction algorithm based on contrast enhancement is proposed.Some natural scene images have low contrast because of blur and illumination. The proposed method which is called local contrast enhancement algorithm based on histogram modification framework is robust against low contrast problem through the comparative experiments. The proposed method can obviously improve the global recall.(3)A sequential character classifier and text region construction based on geometry is proposed.The sequential classifier is constructed by the machine learning algorithm such as Real AdaBoost with Decision Tree and SVM. The classifier makes full use of the character features. The text regions can be effectively constructed from the character regions by determining the distance which is computed based on the weighted sum of the features between characters.We measure the performance of the proposed method on the ICDAR 2013 Robust Reading Competition (Challenge 2:Reading Text in Scene Images) database, the proposed algorithm shows the global precision of 72.9% and the global recall of 87.04%,also including the global F-score of 79.28%.The proposed method achieves expected performance.