This paper describes a new approach to document classification based on visual features alone. Text-based retrieval systems perform poorly on noisy text. We have conducted series of experiments using cosine distance as our similarity measure, selecting varying numbers local interest points per page, and varying numbers of nearest neighbour points in the similarity calculations. We have found that a distance-based measure of similarity outperforms a rank-based measure except when there are few interest points. We show that using visual features substantially outperforms textbased approaches for noisy text, giving average precision in the range 0.4-0.43 in several experiments retrieving scientific papers.

document classification, SIFT