Abstract

In image search re-ranking, a major problem restricting the image retrieval
development is a intent gap, which is a gap between user’s real intent and query/demand
representation, besides well-known semantic gap. In the past, for achieving effective web
image retrieval, classifier space or feature space is explored at a time by researchers. Visual
information and images initial ranks with single feature are only considered in conventional
re-ranking techniques for measuring typicality and similarity in web image retrieval, while
overlooking click-through data influence. For image retrieval, various image features
aggregation shows its effectiveness in recent days. But, uplifting the best features impact for a
specific query image presents a major challenge in computer vision problem. In this paper,
Weight based Multi-Feature Fusion (WMFF) is fused by Improved Artificial Bee Colony
(IABC) for presenting a re-ranking algorithm to retrieve web image. Based on web query,
features are assigned with weights, where different weights are received by different queries in
ranked list. IABC algorithm used to compute weights is a data-driven algorithm and it does
not require any learning. At last, in a web, color and texture features are fused using fusion
and these features are extracted with respective modalities. A Semi-supervised Consensus
Clustering re-ranking with click-based similarity and typicality procedure termed as SCCCST
is used in re-ranking technique. Convolutional Neural Network (CNN) classifier with Multiple
Kernel Learning (CNN-MKL) is used here for performing click-based similarity. Its operation
is depends on selection of click-based triplet’s and a classifier is used for integrating multiple
features into a unified similarity space. The web image search re-ranking performance is
greatly enhanced using proposed technique.