Affiliations: [a] Huazhong University of Science and Technology, Wuhan, China
| [b] Wuhan Yu Ran Intelligent Technology Co., Ltd., Wuhan, China
Correspondence:
[*]
Corresponding author. Qubo Xie, Huazhong University of Science and Technology, Wuhan, China. E-mail: baidu@hust.edu.cn.
Abstract: Recently, many breakthroughs have been achieved in the text detection field; however, printed text detection performance remains unsatisfactory. To address this issue, this paper proposes a refined feature attention based text detection model comprising a feature attention FCNs and text instance segmentation. With the feature attention mechanism, the FCNs model is optimized effectively, which enables the network to learn more precise and accurate features. Therefore, the network can better detect noise-intensive and dense text in medical images. In addition, a centerline-based text region detection algorithm is proposed to process the output of network during text instance segmentation. This algorithm calculates each text region according to the geometric information of the text instance; thus, it is able to process multi-oriented text instances precisely. The proposed model can be trained end-to-end and does not require post-processing operations, which greatly increases detection efficiency. The proposed model achieved excellent results on a medical text image dataset. Compared to existing text detection models, the proposed model demonstrates significantly better performance in terms of F-meatures and detection speed.
Keywords: Printed text detection, attention mechanism, medical text image, image
instance segmentation
