Text detection techniques have become highly mature, but detecting text with arbitrary shapes remains a major challenge in text detection tasks due to geometric encoding limitations of text boxes. In natural scenes, text exhibits a diverse range of shapes, and real-world text is subject to influences such as shooting angles, physical distortions of background objects, and the inherent curvature of the text itself. Merely using rectangular bounding boxes proves insufficient for encompassing irregular text instances. To improve the issue of detecting text with arbitrary shapes, a method that utilizes Fourier transforms in the frequency domain to construct text features is proposed, enabling the reconstruction of predicted boxes by predicting the text’s central line. The predicted text central line not only assists in reconstructing text boxes with complex shapes, but also aids in the subsequent text recognition process through central line correction. The proposed method achieves quite competitive performance on challenging arbitrary-shaped text detection dataset CTW1500, TotalText.