|
The detection of arbitrarily shaped scene text is an extremely challenging task. The existing segmentation-based methods have difficulty in distinguishing adjacent text instances and achieve unsatisfactory results. To solve this problem, a novel weighted segmentation network (WSN) that accurately detects text centers and distinguishes adjacent text instances is proposed in this study. In the WSN, a weighted segmentation map is generated by assigning different weights to each pixel in the text area. Then, we extract features from the weighted segmentation map and detect text centers accurately to distinguish adjacent text instances. To train the detector on the weighted segmentation map, we designed a method for annotation generation based on polygon scaling. Extensive experiments on two benchmark datasets, namely Total-Text and CTW-1500, which comprise highly curved texts in natural scene images demonstrated that our WSN can achieve great performance during segmentation-based methods. |
|
Keywords:Scene text detection; Weighted segmentation; Text centers; Polygon scaling |
|