Comparison of Different Semantic Negative Concepts Selection Methods in SVM Classifier Training for Image Annotation
Shan-Bin Chan, Hayato Yamana · 2013
When SVM is adopted for image annotation, most researchers randomly choose negative sample images for classifier training. Adopting different negative sample image datasets will vary annotation accuracy. This research discusses the accuracy and mean reciprocal rank (MRR) between different negative sample images selection methods. This research adopted ImageNet dataset for positive and negative sample images, and implement SVM for classifiers training. Then we adopted WordNet for building semantic hierarchical tree, and then propose six different negative sample images selection methods. The results show that the accuracy of baseline method (random sampling) is 0.48 and the best proposed method is 0.51.