Curved text revealed: enhancing detection with multidirectional pooling and spatial awareness
Guoxiang Tong, Zhenghao Fu, Taige Dong · Journal of Electronic Imaging · 2025
Existing scene text detection methods are primarily based on deep learning frameworks for pixel-level semantic segmentation to accurately identify and locate text regions in complex scenes. However, when confronted with curved text, these methods often face accuracy challenges due to unclear boundaries and shape complexity. In addition, they have high computational demands, typically requiring significant resources and time. To address these issues, we propose EDMPSA, a curved text detection network that enhances detection with multidirectional pooling and spatial awareness. The network first captures multidirectional features around text regions using a multidirectional pooling module and then expands the model’s receptive field with a spatial awareness module to better understand contextual information. Meanwhile, by combining keypoint shape representation and the keypoint offset regression module, EDMPSA efficiently fits the geometry of curved text, avoiding complex postprocessing. Our ablation studies and comparisons on the Total-Text, CTW-1500, and TD500 datasets show that the model outperforms most state-of-the-art methods for detecting both curved and regular text.