Abstract: An improved self-adaptive method for text area extraction is proposed.With it，the document image is segmented into text area and non-text area firstly.And then，for text area，local features and global features are extracted.The local features include gaps between connected characters，height and width of connected characters，and the global features contain writing style and paragraph features.For non-text area，the key block feature is extracted.After that，the retrieval method combines all the features to improve the accuracy.Meantime，multi-dimensional retrieval structure is introduced to improve the speed.The experiments performed on a large-scale document image database （including 12，024 images） reveal that the method is more efficient than existing ones.

Key words: document image retrieval, text area extraction, paragraph feature, multi-dimensional retrieval structure

摘要： 提出一种改进的自适应文字区域提取算法，将文档图像分割成文字区域和非文字区域。对文字区域提取连通字符间空白、连通字符高度和宽度等局部特征，以及书写样式、段落特征等全局特征；对非文字区域，提取关键块特征。然后利用检索算法将文字区域特征和非文字区域特征结合起来，提高检索的准确性。同时，在检索算法中引入多维数据检索结构，有效地提高检索速度。通过对大规模文档数据库（包含12 024个文档）的检索，表明该算法具有较高的效率，优于现有的一般文档图像检索算法。

关键词: 文档图像检索, 文字区域提取, 段落特征, 多维数据检索结构