Scene Text Detection Based on Bi-directional Feature Pyramid Network

A scene text detection algorithm based on bidirectional feature pyramid network is extracted to address the problem of inaccurate text region detection caused by text of different sizes and shapes in natural scenes. Firstly, the ordinary convolution is replaced by variability convolution in the backbone network, which can better locate the region of interest, and then BIFPN is introduced to remove the unnecessary branches in it as a feature fusion network to fuse the feature information on different scales, and finally differentiable binarisation is used to generate the text region. The proposed method is experimented on the publicly available datasets ICDAR2015, MSRA-TD500, and Total-Text dataset, and its F-value reaches 87.4%, 86.5%, and 85.6%, respectively. The experimental results show that the method has good detection performance on natural scene text.