CONTENT BASED IMAGE RETRIEVAL USING DEEP LEARNING

Abstract

Content-Based Image retrieval(CBIR) is a technique to search and retrieve similar
images from large multimedia databases and an IR system is regarded as efficient if it can
retrieve all the images to meet the user’s needs. There are many advancedmachine-learning
technologies such as deep neural networks(DNN), convolutional neural networks(CNN),
and transfer-learning(TL), which are gaining greater importance in image-related tasks. In
this paper an efficient framework for content-based image retrieval system adapting
transfer-learning on pre-trained CNNs (ResNet18, GoogLeNet, AlexNet) using query-byimage method is proposed, the method explores classification-score descriptors for IR and
employ distance metrics for similarity matching. The framework prescribes transferlearning for efficient retraining of pre-trained CNNs on small datasets chosen from the
Wang database. Thirty-plus experiments are designed for finding optimal values of the
hyper-parameters and exploring the suitability of six popular distance metrics namely
Euclidean, seuclidean, Cityblock, Cosine, Mahalanobis, and Chebychev. After extensive
experimentation, a new efficient framework for CBIR using CNN classification scores is
proposed and the new framework of CBIR achieves the image retrieval accuracy of 99.45%
on natural scene images of 20 classes of the Wang dataset. The experimentation show that
the proposed framework is efficient for content-based image retrieval system