Mask TextSpotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes

Unifying text detection and recognition in an end-to-end training fashion has become a new trend for reading the wild, as these two tasks are highly relevant complementary. In this paper, we investigate problem of scene spotting, which aims at simultaneous natural images. An trainable neural network named Mask TextSpotter is presented. Different from previous spotters that follow pipeline consisting proposal generation sequence-to-sequence network, enjoys simple smooth learning procedure, both can be achieved directly two-dimensional space via semantic segmentation. Further, spatial attention module proposed to enhance performance universality. Benefiting representation on recognition, it easily handles instances irregular shapes, instance, curved text. We evaluate four English datasets one multi-language dataset, achieving consistently superior over state-of-the-art methods tasks. Moreover, further our method separately, significantly outperforms regular recognition.