Research on Tibetan Newspaper Image Layout Reconstruction Based on Dynamic Semantic Injection

To reconstruct outdated Tibetan newspaper images into electronic documents and promote the digital retrieval and preservation of Tibetan information, this paper proposes a Tibetan newspaper image layout reconstruction method based on dynamic semantic injection. The method consists of four key stages: text detection, character recognition, layout analysis, and layout reconstruction. In the text detection stage, the MSF-DBNet model is designed to accurately locate text block positions within the document images. In the character recognition stage, the Mob-CRNN network is innovated to extract Tibetan feature information from the detected regions and efficiently transcribe it into text strings. In the layout analysis stage, the BiF- PicoDet model is adopted to detect and classify document regions, including paragraphs, titles, and images. In the layout reconstruction stage, compared to static preset layout schemes, a dynamic semantic injection approach is proposed to flexibly inject content into corresponding layout areas based on the document's structure and semantic features, thereby effectively adapting to the variable layout designs of newspapers. Experimental results demonstrate that the proposed method achieves excellent performance on the Aba Tibetan Newspaper Dataset (AbaTND) and the PubLayNet public dataset, successfully reconstructing newspaper images into electronic document formats while preserving the visual consistency of the original content.