A Multi-Level Feature Fusion Network For Scene Text Detection with Text Attention Mechanism

To solve the problems of missed text detection and inaccurate text region location in natural scene text detection, a multi-scale and multi-level feature fusion network with attention mechanism is proposed. This method uses the Mask RCNN as the basic framework and improve the backbone ResNet with deformable convolution and rectangular pooling, which extracts multi-level features including global-level, word-level, and character-level, which can extract more comprehensive and richer text feature information. Besides, this paper proposes a region proposal network with text attention mechanism. Experimental results show that the algorithm model effectively extract more useful feature information, significantly improves the recall and accuracy of text detection compared with the current existing methods, and can be applied to actual text detection tasks.