Memory Reduction for Real-Time Document Image Retrieval with a 20 Million Pages Database
Kazutaka Takeda, Koichi Kise, Masakazu Iwamura · 2011
Abstract—We have introduced the three improvements of Locally Likely Arrangement Hashing (LLAH) in ICDAR2011 to reduce a required amount of memory and increase discrimination power of features. In this paper, we show the experimental results which is obtained on a larger-scale database than that utilized for ICDAR2011. From experimental results, we have confirmed that the proposed method realizes 60 % memory reduction and achieves 99.2 % accuracy with 49ms/query processing time for the retrieval of a database of 20 million pages. Keywords-Document image retrieval, Real-time 20 million pages processing, LLAH, Large-scale database I.