CC-DBNet: A Scene Text Detector Combining Collaborative Learning and Cascaded Feature Fusion

In recent years, scene text detection technologies have received more and more attention and have made rapid progress. However, they also face some challenges, such as fracture detection in text instances and the problem of poor robustness of detection models. To address these issues, we propose a scene text detector called CC-DBNet. This detector combines Intra-Instance Collaborative Learning (IICL) and the Cascaded Feature Fusion Module (CFFM) to detect arbitrary-shaped text instances. Specifically, we introduce dilated convolution blocks in IICL, which expand the receptive fields and improve the text feature representation ability. We replace the FPN in DBNet ++ with a CFFM incorporating efficient channel attention (ECA) to utilize features of various scales better, thereby improving the detector's performance and robustness. The results of the experiment demonstrate the superiority of the proposed detector. CC-DBNet achieves 88.1%, 86%, and 88.6% F-measure on three publicly available datasets, ICDAR2015, CTW1500, and MSRA-TD500, respectively, with 0.8%, 0.7%, and 1.4% improvement compared with the baseline DBNet ++, respectively.