A Visual Saliency-Based Approach for Content-Based Image Retrieval

During﻿the﻿past﻿two﻿decades﻿an﻿enormous﻿amount﻿of﻿visual﻿information﻿has﻿been﻿generated;﻿as﻿a﻿result,﻿ content-based﻿image﻿retrieval﻿(CBIR)﻿has﻿received﻿considerable﻿attention.﻿In﻿CBIR﻿the﻿image﻿is﻿used﻿ as﻿a﻿query﻿to﻿find﻿the﻿most﻿similar﻿images.﻿One﻿of﻿the﻿biggest﻿challenges﻿in﻿CBIR﻿system﻿is﻿to﻿fill﻿up﻿ the﻿“semantic﻿gap,”﻿which﻿is﻿the﻿gap﻿between﻿low-level﻿visual﻿features﻿and﻿the﻿high-level﻿semantic﻿ concepts﻿of﻿an﻿image.﻿In﻿this﻿paper,﻿the﻿authors﻿have﻿proposed﻿a﻿saliency-based﻿CBIR﻿system﻿that﻿ utilizes﻿the﻿semantic﻿information﻿of﻿image﻿and﻿users﻿search﻿intention.﻿In﻿the﻿proposed﻿model,﻿firstly﻿ a﻿significant﻿region﻿is﻿identified﻿with﻿the﻿help﻿of﻿method﻿structured﻿matrix﻿decomposition﻿(SMD)﻿ using﻿high-level﻿priors﻿that﻿highlight﻿the﻿prominent﻿area﻿of﻿the﻿image.﻿After﻿that,﻿a﻿two-dimensional﻿ principal﻿component﻿analysis﻿(2DPCA)﻿is﻿used﻿as﻿a﻿feature,﻿which﻿is﻿compact﻿and﻿effectively﻿used﻿ for﻿fast﻿recognition.﻿Experiment﻿results﻿are﻿validated﻿on﻿different﻿image﻿dataset﻿having﻿an﻿extensive﻿ collection﻿of﻿semantic﻿classifications.

This article, published as an Open Access article on November 6, 2020 in the gold Open Access journal, International Journal of Cognitive Informatics and Natural Intelligence (converted to gold Open Access January 1, 2021), is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/)which permits unrestricted use, distribution, and production in any medium, provided the author of the original work and original publication source are properly credited.

Region of Interest Identification
In the proposed approach, an existing saliency detection method known as Structured Matrix Decomposition (SMD) Houwen Peng et al. (2017) is used to detect the region ofinterestintheimage.Inthismethodthelow-levelfeaturesareintegratedwithhighlevel priors to find the salient region.The low-level salient object detection consists of four steps: image abstraction, index tree construction, matrix decomposition and saliencyassignmentHouwenPengetal.(2017).Intheimageabstractionstep,aninput image is partitioned into compact and perceptually homogeneous patches.Each patch Pi is represented by a feature vector fi, and all these feature vectors form the feature matrixasF Thehigh LetAisann-dimensionalunitarycolumnvector.Theideaof2DPCAistoprojectagivenimage X,amxnmatrix,ontoAtogetanm-dimensionalvectorYbythelineartransformation: AuthorscallYtheprojectedfeaturevectorofX.Theprocedureof2DPCAtoobtaintheAvector canbecharacterizedasfollows.
Figure 7 depicts the parametric comparison of different approaches on images categories of Coral10Kdatasetlistedinthetable1.ResultiscomparedwiththeexistingapproachSRBIRand GSD.Itisdemonstratedfromthegraphthatproposedmodelshowsbetterresultsevenincaseof complexbackground.
Figure 9 depicts the parametric comparison of different approaches on images categories of Olivia-2688datasetlistedinthetable3IntheOlivia-2688datasetimagesarecomplexandvisually less different like images of category inside city, street, tall building and highway are very less differentinvisualappearance.ResultiscomparedwiththeexistingapproachSRBIRandGSDand itisanalyzedthatproposedmodelshowsbetterresults.The proposed methodology presented in the paper is able to extract the foreground region effectivelyinthecaseofcomplexbackgroundimage;thebackgroundoftheimageissuppressed successfully.In the technique after identifying the region of interest single feature is taken into considered.Themainaimoftheproposedtechniqueistoenhancetheaccuracyandefficiencyof theQuerybyImageContent(QBIC)applicationwithusingleastfeature.Theexperimentalresultis calculatedontheparametersPrecision,Recall,AccuracyandF-measure.Itisfoundthattheresult oftheproposedQBIRsystemisequivalenttotheretrievalusingmultiplefeatures.

Figure 1 .
Figure 1.Proposed framework for content based image retrieval

Figure 2 .
Figure 2. Algorithm of proposed image retrieval model

Figure 4 .Figure
Figure 4. Some of the sample images from the coral-10K dataset

Figure
Figure 6.Region of Interest (ROI) extraction using the proposed model and SRBIR

Figure 7 .
Figure 7. Graph comparison of PM with GSD and SRBIR on Coral 10 dataset

Figure 8 .
Figure 8. Graph comparison of PM with GSD and SRBIR on GHIM10k dataset

Figure 9 .
Figure 9. Graph comparison of PM with GSD and SRBIR on Olivia-2688 dataset