MS TextSpotter: An Intelligent Instance Segmentation Scheme for Semantic Scene Text Recognition in Asian Social Networks

Text detection and recognition in natural scenes is an im- important task in computer vision. However, most of the texts in natural scenes are curved, and the text background is complex and diverse. In recent years, the text detection and text recognition models proposed have inherent defects, especially in applying many false positives, which usually leads to a decline in text detection and text recognition accuracy. To solve this problem, we propose a text detection and recognition model based on instance segmentation: MS TextSpotter (Mask Scoring TextSpotter), which is based on end-to-end training. First, we design a neural network based on Mask R-CNN. The neural network can achieve accurate text detection and recognition through semantic segmentation, especially for multi-directional and curved text in natural scenes. Second, the network block we designed combines text features and predictive masks to learn the quality of text masks and regresses the intersection ratio of text masks to improve the quality of character masks. The model was tested on ICDAR2013, IC- DAR2015, and Total-Text datasets. Experimental results show that the text detection recall rate increases by 1.4% and 0.2% under the first two datasets, and the experimental results under three different vocabularies- is also show that MS TextSpotter has higher accuracy than other text recognition models and is more suitable for curved text recognition in natural scenes.