Automatic Image Annotation using Minimum Barrier Salient Object Detection and Random Forest

We proposed a new approach of automatic image annotation (AIA) using salient object detection. Salient objects which are used for AIA are generated applying Minimum Barrier Salient Object Detection (MB) method. To improve the quality of AIA, we combined the features from salient objects and entire images by doing a weighted average. The combinations of the features were used to build a random forest (RF) classifier. The classifier was used to produce labels for the testing images. We used Corel 5K as the dataset in this work consisting of 50 categories of images. However, since the saliency approach seeks the dominant object in an image, only 25 groups of images (2500 images) with explicit objects in the image were tested in this study. The result shown that the highest accuracy was obtained when the feature was averaged, using the weight of 2/17 for the salient-object feature and 15/17 for the overall-image feature. Image labeling accuracy reached 79.89% with 95%-CI ranging between 76 to 83%. This proves our proposed approach, with the RF generated using training data in the form of a combined feature, can perform better classification than RF which uses the whole image features or RF that only uses the features of salient object.