Refinement Correction Network for Scene Text Detection

In scene text detection, the accurate capture of underlying detail information and high-level semantic information are crucial for the accuracy and reliability of text detection. To this end, existing models primarily employ deep convolutional networks to extract semantic information from images. However, the multiple convolutions and downsampling operations in network lead to varying degrees of defects in shallow and deep features. To address this issue, this paper proposes the Refinement Correction Network (RCNet). Specifically, in the feature extraction process, constructing a Rough Feature Refinement Module (RFRM) based on the idea of image histogram equalization to restore the texture details of coarse results using underlying features. By modeling high-level features in multiple dimensions, a Clue Feature Correction Module (CFCM) is designed to enhance the semantic relevance of high-level features in spatial and channel positions. Experiments on four benchmark datasets validate the superiority of the proposed model over current technologies.