A new image labeling method based on content-based image retrieval and conditional random field
Xiaofeng Wang, Xiao-Ping Steven Zhang, I. Clarke, Y. Yakubovich · 2009
This paper presents a new image labeling approach that implicitly incorporates top-down information using content-based image retrieval (CBIR) with conditional random field (CRF) model. To reduce the content ambiguities a small content similar training set for CRF labeling is built using retrieved matches from CBIR. To achieve global consistency of image labeling, a novel CRF probabilistic model with a revised global factor is also presented. The proposed method is devised for large labeled databases by learning the top-down content information with CBIR and integrating CBIR retrieval information with the CRF model. The new image labeling model base on CBIR and CRF is compared with the CRF approach without retrieval and demonstrates promising results for floor labeling with Labelme database.