Paper Info
Title
Automatic image tagging as a random walk with priors on the canonical correlation subspace
Abstract
In this paper, we present a graph-based scheme founded on the GCap method of Pan et al. [12] to perform automatic image annotation. Our approach, namely enhanced GCap (EGCap), takes advantage of the canonical correlation analysis technique (CCA) to shorten the semantic gap in the image space and define a new metric in the text space to correlate annotations. As a result, graph linkage errors at the image level are decreased and the consistency of tags output by the system is improved. Besides, we introduce graph link weighting techniques based on inverse document frequency and CCA metric which are proved to enhance the annotation quality. Simple and self-consistent, the present approach achieves image tagging in real time due to the lightweight Local Binary Pattern image features used, the absence of image segmentation, and the reduced size of feature vectors after CCA projection. We test the proposed approach against top-grade state-of-the-art techniques on Corel and Flickr databases, and show the effectiveness of our method in terms of per-word, per-image and processing time performance indicators.
Year
DOI
Venue
2008
10.1145/1460096.1460110
Multimedia Information Retrieval
Keywords
DocType
Citations 
cca projection,random walk,image level,annotation quality,image space,gcap method,present approach,automatic image annotation,cca metric,canonical correlation subspace,image segmentation,feature vector,image retrieval,real time,image features,performance indicator,semantic gap,image annotation,local binary pattern,canonical correlation analysis,canonical correlation,inverse document frequency
Conference
13
PageRank 
References 
Authors
0.57
10
3
Name
Order
Citations
PageRank
Timothée Bailloeul1402.68
Cai-Zhi Zhu21278.82
Yinghui Xu317220.23