Text detection with selective large-kernel and spatial relation

Segmentation-based text detection methods can accurately locate arbitrary-shaped text regions. However, there are two problems to be solved in the current segmentation-based arbitrary-shaped text detection methods:Accurate text detection in images requires extensive background information, and text of different shapes require very different contextual information; Existing methods often fail to incorporate contextual features and fully leverage the features of all scales, thereby ignoring the relationships between feature maps at different stages. Aiming at these problems, this paper proposes an arbitrary-shaped scene text detection method LSRNet. In this network, a selective large-kernel network module is designed to effectively deal with the variation in background information requirement for different text shapes by dynamically adjusting the receptive field, and the receptive field is increased with minimal additional parameters. In addition, a spatial relationship network module using contextual features is proposed to make full use of the connections between feature maps at different stages. Experiments conducted on public datasets show that the proposed method achieves good performance in arbitrary-shaped text detection.