A Multiple Instance Learning Approach to Image Annotation with Saliency Map

Abstract

This paper presents a novel approach to image annotation based on multi-instance learning (MIL) and saliency map. Image Annotation is an automatic process of assigning labels to images so as to obtain semantic retrieval of images. This problem is often ambiguous as a label is given to the whole image while it may only corresponds to a small region in the image. As a result, MIL methods are suitable solutions to resolve the ambiguities during learning. On the other hand, saliency detection aims at detecting foreground/background regions in images. Once we obtain this information, labels and image regions can be aligned better, i.e., foreground labels (background labels) are more sensitive to foreground areas (background areas). Our proposed method, which is based on an ensemble of MIL classifiers from two views (background/foreground), improves annotation performance in comparison to baseline methods that do not exploit saliency information.