원문정보
초록
영어
An image retrieval method based on the combination of visual dictionary and region of saliency was proposed in this paper, which aims to increase the accuracy of image retrieval. The image is divided into sampled blocks and the low-level features are extracted from these image blocks. Then a variety of features vectors are taken as the input vector for learning its corresponding visual dictionary respectively by non-negative sparse coding. Spatial information is added into the sparse representations of image by proposing the saliency polling method, and the similarity measure between sparse representation vectors is defined as SED (Squared Euclidean Distance), which considering the same nonzero entries and Euclidean distance of vectors at the same time. Results of experiment carried on Corel and Caltech datasets showed that this method can effectively improve the accuracy of image retrieval compared with the methods of single visual dictionary.
목차
1. Introduction
2. Non-Negative Sparse Dictionary
3. Multiple Visual Dictionary of Saliency Weighted Image Retrieval
3.1. Saliency Pooling
3.2. Multiple Visual Dictionary Image Representation
3.3. Similarity Measurement
4. Experiment
4.1. Sub-Block Division and Feature Extraction
4.2. Experiment Settings
4.3. Corel10K Dataset
4.4. Caltech Dataset
5. Conclusion
Acknowledgements
References