Content-based similar document image retrieval using fusion of CNN features

摘要

Rapid increase of digitized document give birth to high demand of documentimage retrieval. While conventional document image retrieval approaches dependon complex OCR-based text recognition and text similarity detection, this paperproposes a new content-based approach, in which more attention is paid tofeatures extraction and fusion. In the proposed approach, multiple features ofdocument images are extracted by different CNN models. After that, theextracted CNN features are reduced and fused into weighted average feature.Finally, the document images are ranked based on feature similarity to aprovided query image. Experimental procedure is performed on a group ofdocument images that transformed from academic papers, which contain bothEnglish and Chinese document, the results show that the proposed approach hasgood ability to retrieve document images with similar text content, and thefusion of CNN features can effectively improve the retrieval accuracy.