Image annotation using label propagation algorithm

An approach to image annotation is proposed. Generally, the relation between visual characteristics and the annotation label is estimated from the annotated corpus and is used to predict label for new test image. Unfortunately, when limited number of images are annotated, with possible multiple labels per image, this relation cannot be reliably estimated. Moreover, the common approach cannot take advantage of available un-annotated images which are easier to gather. This work applies a label propagation algorithm to assign the label posterior probability to images using information from available un-annotated images in semi-supervised manner. Experimental results show that the performance of this model is encouraging.