Abstract:

We propose a novel classifier for the multilabel image annotation task based on an improved SVM. We firstly define a histogram intersection distance for the SVM kernel function. Then, the original SVM output result is transformed to the distance between a given sample and the hyperplane. Additionally, a feature selection method is developed for our model, and we choose those visual features with small correlations between them to establish the SVM based classifier. Furthermore, on account of the uneven distribution of different image categories, we also introduce a probability weighted strategy in our SVM model. Experiments on ImageCLEF dataset not only confirm the effectiveness of the proposed model, but also show that the proposed feature selection method is very suitable for the classifier. Compared with the traditional classifiers, our method obtains the optimal results, and is competitive to the state-of-the-art methods.

Key words: SVM;image annotation;multiple classifiers;feature selection