Text kernel calculation for arbitrary shape text detection

With the speedy progress of deep learning, text detection has received progressively increasing attention and considerable progress. The current mainstream approaches are usually based on instance segmentation to obtain the label of whether the pixel is text, as this can cope with arbitrary-shaped text. However, pixel-based prediction usually leads to overlapping neighboring texts, resulting in misdetection. To mitigate the above problems, we propose an approach to calculate text kernels and determine the attribution of boundary pixels. This way, all texts are labeled uniformly, facilitating model learning and effectively separating adherent texts. In addition, to cope with the complex and variable background of the text, we propose a practical feature enhancement module to handle it. The proposed module can explore different levels of features to represent text information of diverse sizes. Compared with current advanced algorithms, our method is competitive, which achieves the F1-measure of 87.3, 88.0, 82.8, 85.7, and 90.0% on the ICDAR2015, MSRA-TD500, CTW1500, Total-Text, and ICDAR2013 datasets, respectively.