Scene Uyghur Text Detection Based on Adaptive Feature Fusion

Dong Wang, Elham Eli, Alimjan Aysa, Xuebin Xu, Hornisa Mamat, Kurban Ubul

Abstract: In recent years, with the development of deep learning, text detection research has achieved good research results. However, there is still relatively little research on the detection of Uyghur text in natural scenes. Therefore, this paper proposes a scene Uyghur text detection model based on adaptive feature fusion for scene Uyghur texts with special writing styles, complex and variable backgrounds, and different text scales. First, a normalization-based attention module is introduced into the feature extraction network to enhance text features while suppressing background noise. Second, in order to better extract the features of multi-scale text, this paper adds the proposed adaptive feature fusion module in the feature fusion stage. The efficient fusion of text features at different scales is realized by adaptively adjusting the weights between features at different levels. Finally, experiments on the Scene Uyghur text dataset and the ICDAR2015 dataset show the effectiveness and robustness of the proposed method in this paper.