Composite Descriptors and Deep Features Based Visual Phrase for Image Retrieval

摘要

Local descriptors are very effective features in bag-of-visual-words (BoW) and vector of locally aggregated descriptors (VALD) models for image retrieval. Different kinds of local descriptors represent different visual content. We recognize that spatial contextual information play an important role in image matching, image retrieval and image recognition. Therefore, to explore efficient features, firstly, a new local composite descriptor is proposed, which combines the advantages of SURF and color name (CN) information. Then, VLAD method is used to encode the proposed composite descriptors to a vector. Third, local deep features are extracted and fused with the encoded vector in the image block. Finally, to implement efficient retrieval system, a novel image retrieval framework is organized a novel image retrieval framework is organized based on the proposed feature fusion strategies. The proposed methods areis verified on three benchmark datasets, i.e., Holidays, Oxford5k and Ukbench. Experimental results show that our methods achieves good performance. Eespecially, the mAP and N-S score achieve 0.8281 and 3.5498 on Holidays and Ukbench datasets, respectively.