Combining firefly algorithm and Bayesian classifier: new direction for automatic multilabel image annotation

摘要

As digital images are increasing exponentially; it is very attractive to develop more effective machine learning frameworks for automatic image annotation. In order to address the most prominent issues (huge inter-concept visual similarity and huge intra-concept visual diversity) more effectively, an inter-related non-parametric Bayesian classifier training framework to support multi-label image annotation is developed. For this purpose, an image is viewed as a bag, and its instances are the over-segmented regions within it found automatically with an adopted Otsu's method segmentation algorithm. Here firefly algorithm (FA) is utilised to enhance Otsu's method in the direction of finding optimal multilevel thresholds using the maximum variance intra-clusters. FA has high convergence speed and less computation rate as compared with some evolutionary algorithms. By generating blobs, the extracted features for segmented regions, the concepts which are learned by the classifier tend to relate textually to the words which occur most often in the data and visually to the easiest to recognise segments. This allowing the opportunity to assign a word to each object (localised labelling). Extensive experiments on Corel benchmark image datasets will validate the effectiveness of the proposed solution to multi-label image annotation and label ranking problem.