Hierarchical Feature Fusion With Text Attention For Multi-scale Text Detection
Chao Liu, Yuexian Zou, Wenjie Guan · 2018
Scene text detection for practical application remains challenging since it is a typical multi-scale object detection task with complex and varying condition. It is noted that the single shot detector (SSD) has shown predominance in object detection among deep learning-based methods but has limited capability in handling multi-scale text detection (MTD) task. In this paper, we strive to improve the performance of MTD task under SSD framework. Specifically, a novel single-shot word-level text detector is proposed. First, to extract the features keeping text-related information for multi-scale text objects, a hierarchical feature fusion module is designed to capture multi-scale inception features and multi-level features with semantic information. Second, to suppress background disturbance in the feature map, a text attention module is developed to learn the location information of the texts existing in the image, which helps to accurately inferring the words, especially when they are of extremely small size. Experimental results on three public word-level texts datasets demonstrate the effectiveness of our proposed method which achieves F-measure 0.89 and 0.79 on the ICADR2013 and ICDAR2015 respectively, and achieve highest accuracy of 0.87 on the SVT.