AncientGlyphNet: an advanced deep learning framework for detecting ancient Chinese characters in complex scene

Detecting ancient Chinese characters in various media, including stone inscriptions, calligraphy, and couplets, is challenging due to the complex backgrounds and diverse styles. This study proposes an advanced deep-learning framework for detecting ancient Chinese characters in complex scenes to improve detection accuracy. First, the framework introduces an Ancient Character Haar Wavelet Transform downsampling block (ACHaar), effectively reducing feature maps’ spatial resolution while preserving key ancient character features. Second, a Glyph Focus Module (GFM) is introduced, utilizing attention mechanisms to enhance the processing of deep semantic information and generating ancient character feature maps that emphasize horizontal and vertical features through a four-path parallel strategy. Third, a Character Contour Refinement Layer (CCRL) is incorporated to sharpen the edges of characters. Additionally, to train and validate the model, a dedicated dataset was constructed, named Huzhou University-Ancient Chinese Character Dataset for Complex Scenes (HUSAM-SinoCDCS), comprising images of stone inscriptions, calligraphy, and couplets. Experimental results demonstrated that the proposed method outperforms previous text detection methods on the HUSAM-SinoCDCS dataset, with accuracy improved by 1.36–92.84%, recall improved by 2.24–85.61%, and F1 score improved by 1.84–89.08%. This research contributes to digitizing ancient Chinese character artifacts and literature, promoting the inheritance and dissemination of traditional Chinese character culture. The source code and the HUSAM-SinoCDCS dataset can be accessed at https://github.com/youngbbi/AncientGlyphNet and https://github.com/youngbbi/HUSAM-SinoCDCS.