Conceptual text region network: Cognition-inspired accurate scene text detection

Segmentation-based methods are widely used for scene text detection due to their superiority in describing arbitrary-shaped instances. However, two major problems still exist: (1) current label generation techniques mostly empirical and lack theoretical support, discouraging elaborate design; (2) as a result, most rely heavily on kernel segmentation which is unstable requires deliberate tuning. To address these challenges, we propose human cognition-inspired framework, termed, Conceptual Text Region Network (CTRNet). The framework utilizes Regions (CTRs), class of cognition-based tools inheriting good mathematical properties, allowing sophisticated design. Another component CTRNet an inference pipeline that, with the help CTRs, completely omits need segmentation. Compared previous segmentation-based methods, our approach not only more interpretable but also accurate. Experiment results show that achieves state-of-the-art performance benchmark CTW1500, Total-Text, MSRA-TD500, ICDAR 2015 datasets, yielding gains up 2.0%. Notably, best knowledge, among first models achieve F-measures higher than 85.0% all four benchmarks, demonstrating remarkable consistency stability.