Automatic image annotation using CMRM with scene information

Abstract

Searching of digital images in a disorganized image collection is a challenging problem. One step of image searching is automatic image annotation. Automatic image annotation refers to the process of automatically assigning relevant text keywords to any given image, reflecting its content. In the past decade many automatic image annotation methods have been proposed and achieved promising result. However, annotation prediction from the methods is still far from accurate. To tackle this problem, in this paper we propose an automatic annotation method using relevance model and scene information. CMRM is one of automatic image annotation method based on relevance model approach. CMRM method assumes that regions in an image can be described using a small vocabulary of blobs. Blobs are generated from segmentation, feature extraction, and clustering. Given a training set of images with annotations, this method predicts the probability of generating a word given the blobs in an image. To improve annotation prediction accuracy of CMRM, in this paper we utilize scene information incorporate with CMRM. Our proposed method is called scene-CMRM. Global image region can be represented by features which indicate type of scene shown in the image. Thus, annotation prediction of CMRM could be more accurate based on that scene type. Our experiments showed that, the methods provides prediction with better precision than CMRM does, where precision represents the percentage of words that is correctly predicted.