DKMI: Diversification of Web Image Search Using Knowledge Centric Machine Intelligence

Abstract

Web Image Recommendation is quite important in the present-day owing to the large scale of the multimedia content on the World Wide Web (WWW) specifically images. Recommendation of the images that are highly pertinent to the query with diversified yet relevant query results is a challenge. In this paper the DKMI framework for web image recommendation has been proposed which is mainly focused on ontology alignment and knowledge pool derivation using standard crowd-sourced knowledge stores like Wikipedia and DBpedia. Apart from this the DKMI model encompasses differential classification of the same dataset using the GRU and SVM, which are two distinct differential classifiers at two different levels. GRU being a Deep Learning classifier and the SVM being a Machine Learning classifier, enhances the heterogeneity and diversity in the results. Semantic similarity computation using Cosine Similarity, PMI and SOC-PMI at several phases ensures strong relevance computation in the model. The DKMI model yields overall Precision of 97.62% with an accuracy of 98.36% along with the lowest FDR score of 0.03 and is much better than the other models that are considered to be the baseline models.