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

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