Text Detection by Jointly Learning Character and Word Regions

Text detection in natural scenes has developed significantly in recent years. Segmentation-based methods are widely used for text detection because they are robust to detect text of any shape. However, most state-of-the-art works are limited to word/line level detection as character-level data annotation is too expensive. Considering the close connection between characters and words, we propose a detector containing four different headers: Gaussian map, offset map, mask map, and centerline map, to obtain word-level and character-level prediction results simultaneously. Besides, we design a weakly supervised method to fully use the word-level labels of the real dataset to generate character-level pseudo-labels for training We perform rigorous experiments on multiple benchmark datasets. Results demonstrate that our method achieves state-of-the-art results. Specifically, we achieve an F-measure of 85.2 on the dataset CTW1500, which is 1.3% higher than the state-of-the-art methods.