High Performance EMID Framework for Automatic Image Annotation and Retrieval

Abstract

With the enormous growth of digital cameras, the computer society has witnessed great significance and challenge in automatic image annotation as a current research field. Automatic image annotation is an efficient method to resolve the problem of “Semantic Gap”. Image auto-annotation is a challenging problem when a label is provided for the entire training image only instead of the object region. To eliminate labeling ambiguity, image categorization and object localization should be performed simultaneously. Discriminative Multi-Instance Multi-label Learning (MIML) framework can be used for this work by considering each image as a bag and sub-windows in the image as instances. With standard traditional approaches, selecting only one positive sub-window per positive bag may reduce the search space for global optimum; meanwhile, selecting all temporal positive sub-windows may add noise into learning. Hence to solve this issue we propose a new approach which selects a subset of sub-windows per positive bag. Our Proposed Enhanced Multi Instance Differentiation (EMID) algorithm is able to acquire the correlation among instances, correlation among labels, and correlation between instances and labels concurrently, and provides a very good representation and learning prospective. The obtained experimental results obtained reveal that our approach performs extensively better than the existing discriminative MIML approaches and traditional categorization approaches.