Bi-directional Feature Fusion for Fast and Accurate Text Detection of Arbitrary Shapes

Existing segmentation based methods have problems, such as the difficulty in distinguishing adjacent text areas and the low efficiency of model detection caused by the complex steps in the post-processing stage. In order to solve this problem, this article proposes a novel scene text detection model based on fully convolutional network, which can solve the problem that adjacent texts are difficult to distinguish in existing methods and improve the detection speed of the model. First, it constructs a feature extractor to extract multi-scale feature map from the input image. Secondly, the bidirectional feature fusion module is used to fuse the semantic information of the two parallel branches and promote the joint optimization of the two branches. It then effectively differentiates adjacent texts by predicting both a reduced text area map and a full text area map in parallel. The former can guarantee the distinction between different text instances, while the latter can effectively guide the network optimization. Finally, in order to improve the speed of text detection, it proposes a fast and effective post-processing algorithm to generate text boundary boxes. The experimental results show that: on relative datasets, the method proposed in this article achieves the best performance, and improves the F-measure index by 1.0% at most compared with the current best method, and can achieve near-real-time speed, which proves fully the effectiveness and high efficiency of the method.