Yasasvy Tadepalli|Meenakshi Kollati |Swaraja Kuraparthi |Padmavathi Kora |Anil Kumar Budati |Lakshmi Kala Pampana

1Department of ECE,GRIET,Hyderabad,India

2Department of ECE,VNR Vignana Jyothi Institute of Engineering and Technology,Hyderabad,India

Abstract Rapid growth in the transfer of multimedia information over the Internet requires algorithms to retrieve a queried image from large image database repositories.The proposed content-based image retrieval (CBIR) uses Gaussian-Hermite moments as the low-level features.Later these features are compressed with principal component analysis.The compressed feature set is multiplied with the weight matrix array,which has the same size as the feature vector.Hybrid firefly and grey wolf optimization(FAGWO)is used to prevent the premature convergence of optimization in the firefly algorithm.The retrieval of images in CBIR is carried out in an OpenCV python environment with K-nearest neighbours and random forest algorithm classifiers.The fitness function for FAGWO is the accuracy of the classifier.The FAGWO algorithm derives the optimum weights from a randomly generated initial population.When these optimized weights are applied,the proposed algorithm shows better precision/recall and efficiency than other techniques such as exact legendre moments,Region-based image retrieval,K-means clustering and Color descriptor waveletbased texture descriptor retrieval technique.In terms of optimization,hybrid FAGWO outperformed various optimization techniques (when used alone) like Particle Swarm Optmization,Genetic Algorithm,Grey-Wolf Optimization and FireFly algorithm.

1|INTRODUCTION

Digital image acquisition in recent years has increased exponentially from rapid advances in internet and cellular communication,a steady decline in the price of memory,and increased access to technology for everyone.In various applications such as medicine,advertising,education,and mass communication,there is a need for the efficient search and retrieval of images similar to the queried image.These aspects motivated the development of content-based image retrieval (CBIR) to retrieve images relevant to the query[1].In CBIR,the features are collected from an image database using preprocessing and feature-extraction steps.The objective of image preprocessing is to suppress unwanted noise and improve image quality in subsequent steps.Preprocessing steps include denoising [2],Gaussian low-pass filtering(GLPF)[3],and anisotropic distortion [4].The features of the queried image are compared with the features of images in the database using a similarity measure.The motivation of CBIR is to retrieve images similar to the image queried.However,when images in the database vary in their translation,rotation,shape,size,or image-related features,efficient image retrieval becomes difficult.

Moments-based feature extraction is carried out for CBIR.Moments have already been used in the literature for edge detection [5],object recognition [6],facial recognition [7],and other realms of image processing and computer vision.Moments are scalar quantities used to characterize a function and derive its features[8].Directly speaking,they are projections onto a polynomial base.There are two types of moments,namely geometrical and orthogonal moments,that differ in terms of stability and computational capabilities,especially in the discrete domain.Powers are used in the computation of geometrical moments,and for small and large values of the exponent,computed values of the moment increase rapidly as order increases.On the other hand,in orthogonal moments,features are captured with more precision due to the representation of moments in recurrent relation when compared with the standard powers used in geometrical moments.Orthogonal Gaussian-Hermite moments (GHMs) are used as features in the proposed algorithm.The orthogonal moment base functions of different orders exhibit different numbers of zero crossings and very different shapes[9].Hence,orthogonal moments represent better distinct image features based on different modes.These discriminative features are useful in computer vision tasks.There are several orthogonal moments,such as exact Legendre moments and Tchebichef moments.Compared with these two moments,GHMs are powerful and it is mainly due to the zero-crossings function of the GHMs which are distributed more evenly than those of other orthogonal moments such as Legendre and discrete Tchebichef moments [10].GHMs are more immune to noise and avoid the artifacts introduced by the discontinuity window function.In addition,GHMs are translation and rotation invariant.These factors motivated us to use GHMs in the proposed CBIR framework.

The proposed CBIR algorithm's retrieval performance is examined by conducting experimentation on the Corel image database,which contains 10 different categories,as shown in Figure 1.The next step is the feature compression technique using principal component analysis (PCA) [11].This step removes irrelevant information,and thus only discriminant features would be obtained for classification.Discriminant GHM features are obtained with PCA.The compressed GHM features are multiplied by weights to obtain control for improving retrieval efficiency.The optimization algorithms seek to find the best available solution in a given search space [12].Reference [13] developed a CBIR framework using exact Legendre Moments,HVS color quantization with DC coefficients and statistical properties such as variance,mean,and skew of the conjugate symmetric sequency complex Hadamard transform.The authors used two optimization techniques,differential evolution and the firefly algorithm (FA),for feature compression.FA was used to derive optimal weights from the random initial population.The algorithm's major drawback was a cascade of optimization algorithms used for feature compression and feature retrieval that drastically increased computational complexity.In CBIR evaluation,the best available solution is to obtain improved retrieval efficiency.The CBIR system's retrieval efficiency is improved by applying optimal weights to the proposed algorithm's feature vectors[13].These optimal weights are generated from a random initial population of hybrid combinations of two popular nature-inspired optimization algorithms,FA [14,15] and grey wolf optimization(GWO) [16],known in combination as FAGWO.Nonoptimized algorithm performance is low in the case of multimodel databases like Corel 1K.Optimum feature selection increases retrieval efficiency and is achieved with the optimization algorithm.FA has the disadvantage of becoming stuck at local maxima.GWO and FireFly (FF) algorithm hybridization ensures optimal weight selection and maintains the trade-off between the degree of exploitation and exploration in the search space.The exploration and exploitation of the optimization are well balanced in the proposed hybrid FAGWO algorithm.In Section 2,prior works on CBIR,GHM,FF,and GWO are put forward.In Section 3,a concise presentation of orthogonal GHMs and their approximation computation is given.The algorithmic steps of the FAGWO algorithm are described in Section 3,and process implementation is discussed in Section 4.Experimental results are presented in Section 5,and finally,the conclusion is given in Section 6.

FIGURE 1 Different Corel 1K dataset objects

2|LITERATURE REVIEW

CBIR finds its applications in medical image analysis,remote sensing,surveillance,security,textile industries etc.Several techniques already exist in the literature for CBIR.Through virtual posing,IBM has developed query by image content[17].The queries are refined based on various image properties such as color,texture,user-constructed sketches,and drawings[17].Feature vectors formed from geometric shapes,colors,and textures are used to distinguish one class of objects from another in CBIR.Reference[18]proposed a CBIR framework based on color moments considering the mean,variance,and skewness of three color planes.The color distributions were later classified using quantized histogram statistical texture features combined with distance metrics such as Euclidean and Mahalanobis distances.However,this method is highly granular,as it carries out a detailed comparison of histogram bins of queries and images in the database.Another drawback is that histogram bins are highly susceptible to noise.Computational complexity is another limitation of this method.Texture variations in different images can distinguish image features,especially when images contain different textured regions.The texture features obtained from discrete wavelet transform combined with those for color are used to classify textures [19].Although this scheme works well on small texture databases,it cannot support geometric invariance.Reference [20] utilized color histogram and texture measures to obtained a CBIR system.Texture features included entropy,smoothness and uniformity.The experiment was performed on a Corel database with Euclidean distance measurement as the retrieval mechanism.Reference [21] presented a CBIR system with color histogram,color moment,co-occurrence matrices,and wavelet moments.A K-means clustering algorithm is employed in the CBIR to form the initial clusters.The fitness function,which was optimized by the particle swarm optimization (PSO) technique,minimized the distance from a cluster point to the cluster centre to improve retrieval efficiency.PSO-based CBIR schemes provide optimum results,yet they suffer from high computational time.Reference [22] incorporates a CBIR framework exploiting interactive genetic algorithms (IGAs) and support vector machine(SVM)to learn the user perception.In this scheme,a SVM classifier is trained using the user feedback.The unlabelled images which are classified as most relevant by the SVM classifier,are added to the IGA training set.However,the main drawback of CBIR systems employing genetic algorithm(GA) optimization is that they suffer from convergence problems.that is,the problem of degeneracy.

The famous GWO technique is used in many classification algorithms.GWO was combined with a bag of features for facial recognition [23].Facial features were obtained using the Speeded up Robust Features descriptor.Optimal features were selected by K-means clustering combined with GWO.It has been proved that GWO produces better convergence precision for standard benchmark problems than other meta-heuristic techniques.Another good optimization technique is FA.Yang first developed FA in 2007[24].Reference[15]proposed an adaptive CBIR framework exploiting relevance feedback and Gaussian FA.A firefly alone has a high probability of being trapped in local maxima.The FA uses the principles of attractiveness,randomization,and absorption.Random walk is initialized for each firefly and the Levy flight function propels the fireflies' movement in isometric random directions.Gaussian distribution was used for random walk propagation in CBIR [25].It has been demonstrated that the FF with clustering algorithm outperformed the GA and PSO.

Deep neural networks made inroads in CBIR.Series of convolution layers extract the low-level and high-level features from images.High-level features represent the semantics of the data,which in turn provide greater robustness to intra-class variability.Low-level hallmarks are rich in texture and other features that are unique to each image dataset.Low-level features of CNN require a massive amount of dataset with minor variations within each dataset.Hence,careful selection of the dataset images [26] is needed because mis-classification might occur.In addition,CNNs are computationally intensive and require resource-intensive hardware to run.

Moreover,the techniques used for extracting the information from an image are referred to as image analysis [27].Autonomous segmentation is a very critical step in image analysis.The eventual success or failure of image analysis depends on segmentation.The strategies used in image segmentation are based on two attributes of grey values,discontinuity and similarity.The principal focus in similarity is thresholding,and in most thresholding methods,ground truth image is analyzed by human subjects based on visual inspection.This method is error-prone,as it relies on human observer scores.Reference [28] rectified this by proposing a novel approach to automatically generate ground truth for image binarization by consensus of different thresholding methods.The authors concluded that F-measure,modified Hausdorff distance,and edge mismatch error are nonredundant and distinct indicators and are useful in evaluating the performance of the algorithm on image databases.

Very recently,CBIR has been effectively used in material identification and medical imagery.The features used in the material database are grey level co-occurrence matrix(GLCM)and artificial bee colony(ABC)[29].The texture features were extracted using GLCM and ABC and were used for classification and retrieval of the images.Histopathology image segmentation was carried out using region extraction and curve fitting to extract the nuclei [30].

With tremendous improvements in wireless technology,the transmission of images sent through wireless networks has increased to a huge extent.Recently,researchers are seeking to improve the transmission capacity of wireless channels.The analog/digital communication system consists of an encoder,transmitter,channel,receiver,and decoder.Usually,an encoder is any system that incorporates efficient image compression[31].Feature reduction techniques such as GHMs and other techniques can help in transforming and reducing the size of the image.Orthogonal frequency division multiplexing(OFDM) is the most used channel waveform for image transmission in both analog and digital systems [32].Discrete cosine transform,wavelet coding,and embedded block coding along with optimal truncation and scaling were used for image compression in medical images [33].OFDM was used to transmit the image using multicarrier code [34].Image transmission happens in packets,but as the image’s size increases,more energy is required to transmit the images.Energy is depleted when the image is transmitted.To prevent this,an image transmission scheme using a multihop was proposed based on LEACH.This technique eliminates the energy hole problem when the cluster head node compresses the images[35,36].3-D image processing is gaining significance,as much information is captured in 3-D images.A patch map was generated and used for error detection on the depth map[37].This patch map not only reduced the transmission error but also enhanced the quality of the 3-D image.

Much of the information can be reduced if an effective feature vector is generated from the images.Shape is an important low-level feature of an image that helps in identifying significant image regions.Without the help of visual cues,shapes can recognize variations in the image effectively [38].Contour-based and region-based shape descriptors are the two important types of shape descriptors.’Contour’deals with only the boundary,whereas in a region-based descriptor,all the information related to the texture and shape is obtained from the image [39].Fourier descriptors,polygonal approximations,curvature scale space,and shape signatures are examples of contour-based shape descriptors.Moment invariants form a part of region-based descriptors [40].‘Invariant’ features are insensitive to the various deformations in the image.Deformations are rotation variance,scale variance etc.If the polynomial base is orthogonal,the moments are termed orthogonal moments.Examples of orthogonal moments are Zernike moments (ZMs),exact Legendre moments (ELMs)and GHMs[40].ZMs are effective in shape recognition but are very computationally intensive.The coefficients of ZM are mapped to a unit circle,which is a limitation while retrieving greyscale images.When the pixels that fall outside the unit circle are omitted,an error is introduced [40,41].This motivated many researchers to experiment on ELMs and GHMs.A model for computing ELMs is proposed,and it eliminates the need for numerical approximation.A fast and accurate method improved the performance of ELMs.It was [9] who reported that ELMs could represent an image with a minimum amount of information redundancy.ELMs are invariant to translation and scaling.They show optimum performance in terms of retrieval efficiency and computational intensity.The main drawback of the ELMs is rotation sensitivity.GHMs[41]blend the advantages of ELMs [9] and ZMs,thus providing computational simplicity and performance richness in terms of Image retrieval.The introduction of GHMs and their application is first found in[9].Later on,the moments were used in fingerprint recognition and classification [43],3-D face recognition [44],SAR image segmentation [45],and the intelligent license plate recognition system [46].Drawing motivation from the performance of the GHMs,CBIR using GHMs is taken up in this paper.It is postulated that higher-order Gaussian moments in discrete and continuous implementations are good at global feature representation and image reconstruction [9].Chandra Mohan et al.[47] demonstrated a fast way to retrieve the Images using GHMs and SVM.In this paper,random forest (RF) and K-nearest neighbours (KNNs)were used as classifiers.RF is intrinsically suited for multiclass problems consisting of a mixture of numerical and categorical features.Less parameterization makes it easy to program and the nature of RF makes them less likely to overfit.KNN is used in this project because many project classes have less linear separability.The KNN algorithm assumes that similar things exist close.This proximity is calculated using the distance metrics.The use of KNN has been both boon and bane in this paper.Without proper optimization,the KNN technique was found to overfit the data points.The same was the case with RF.The Corel database contains images with different backgrounds,different numbers of objects,and different orientations.Non-optimization algorithms solve such multimodal problems poorly.The optimization algorithm must reach the global maxima in search space to increase retrieval efficiency.An optimizer helps to determine the global maxima by selecting the best weight matrix for the feature vector.FF has the disadvantage of becoming struck at local maxima.GWO has better exploration and exploitation tendencies.Hence,a hybrid optimizer combining the FA and GWO technique,FAGWO,is used in the proposed CBIR framework.

3|PRELIMINARIES

In this section,Gaussian-Hermite moments,classification algorithms,namely,KNN and RF,and the optimization algorithm,namely,FAGWO,are discussed.

3.1|Gaussian-Hermite moments

Gaussian-Hermite moments are a set of moments formed using Gaussian-Hermite polynomials as basis functions.The Hermite and Gaussian polynomials are described as follows:

The Hermite polynomials ofithorder moment are defined as

The relation of orthogonality is

δi,jis known as knocker's delta.The orthogonality condition simplifies the reconstruction of the original function from the generated moments.The main reason of introducing orthogonality in Gaussian-Hermite polynomials are stable and fast numerical implementation,avoidance of high dynamic range of moment values that may lead to loss of precision due to overflow or underflow,and a higher robustness to random noise.The Hermite polynomials are orthogonal in the interval of（-∞，∞）[8] and therefore not well suited to imageprocessing tasks.To suit the moments for image processing applications,Wu and Shen [34] proposed orthogonal GHMsGHMij（x，y，f（x，y））as given in Equation (3):

that depend on image coordinates (x,y).Theg（u，v，σ）,a Gaussian function,is defined as

whereσis its standard deviation.Hermite polynomials are computed quickly using recurrent relation used in Equation (5):

Generally,GHMs are defined in the continuous domain.In the discrete domain,GHMs are fast and stable due to computation with recurrent relations.High dynamic range in moment values is prevented by confining GHMs to a narrow interval of [-1，1].

The following equation is a convenient method for discrete implementation of the moment computation:

The number of GHM features obtained for ordernisn2.These features are compressed tozfeatures using PCA.Forzfeatures,zweight matrices are obtained.The weighed features will be used to train machine learning algorithms for classification.KNN and RF classifiers are used in this paper.

3.2|K-nearest neighbours algorithm

KNN [48] is a supervised machine learning technique that classifies objects based on the closest set of points in a vector space containing features.KNN is termed‘lazy learning’where the function works only locally and the computation is carried out until the classification procedure is completed.A simple technique,KNN is widely used when the spatial distribution of the data is unknown.Each data point goes into one of the K clusters by taking the nearest neighbour and it is denoted by factor ‘K’.Efficiency in classification is improved when the optimum ‘K’ value is chosen.

3.3|Random forest classifier

RF [49] employs an ensemble method for classification and regression.Classification is carried out in RF by constructing a multitude of trees at the training phase,resulting in a class with the majority vote.In RF,multiple trees are grown and discrimination between the classes is achieved by projecting the training data on a stochastically chosen vector subspace before fitting the model at each tree or node,as shown in Figure 2.The decision at each node is not selected deterministically but randomly.Bagging provides flexibility for each model in the ensemble to give each tree equal weights.To promote model variance,bagging trains each model in the ensemble using a randomly drawn subset of the training set.Therefore,the RF algorithm combines random decision trees with bagging to achieve very high classification efficiency.The RF produces classification results comparable to other classifiers and has the advantages of not overfitting,withstanding more noise and ease of measuring out-of-bag error.

FIGURE 2 Random forest classifier with majority vote

3.4|Grey wolf optimization

GWO [16,23] is a meta-heuristic and biologically inspired technique inspired by wolves' hunting behavior,as they move and hunt in packs.Packs hunt for potential prey,and once they find it,they encircle and attack it.A leader would be present in the pack and all the other wolves in the lower rung of hierarchy obey the leader's instructions.The same principle is applied in the GWO technique.There are four types of variables designated as four types of wolves which are alpha (α),beta (β),delta (δ) and omega (ω).The hunting is guided byα,β,δrespectively,with theα,as the leader.ωfollows them in the hunting process.The three main hunting steps,namely searching,encircling,and attacking prey,are implemented in GWO using the following mathematical equations:

wherexpis the position vector of the prey andxindicates the grey wolf's position vector andtis the current iteration.AandCare coefficient vectors,AandCvectors can be computed using Equations (9) and (10):

where components ofaare linearly decreased from 2 to 0,throughout iterations,r1andr2are random vectors in [0,1].αis the best solution.βandδare the next best solutions.These solutions are saved and other agents are forced to update their position according to the best position.

To update the grey wolves position,the following equations are used:

whereC1,C2andC3are computed using Equation (10):

whereXα,Xβ,andXγare the three best solutions at iteration t,A1,A2,andA3computed using Equation (9),Dα,Dβ,andDγdefined in Equation(11),andX(t+1)indicates the position of wolves in the next iteration.

The hunting process is ceased when the prey halts moving and wolves jolt an attack and can be done mathematically by decreasing the value of a during iterations for controlling exploration and exploitation of GWO.

3.5|Firefly and grey wolf optimization

The firefly and grey wolf optimization (FAGWO) hybrid optimization algorithm integrates GWO into FA,one of the nature-inspired swarm intelligence-based optimization algorithms that rely on fireflies' flashing light.The FA computes the optimum position of the particles based on the brightness.The fireflies are unisexual and therefore,any firefly can attract another firefly,disregarding their sex.The firefly's attractiveness is directly proportional to one’s intensity and attraction between them curtails as the distance increases.In a firefly pair,the brighter one attracts the firefly with lesser brightness and the intensity falls when their mutual distance increases.The firefly's movement is unpredictable and random if no firefly is found brighter than a given firefly.

Two vital issues considered in the FA are the variation of light intensity and formulation of the attractiveness.The FA assumes that a firefly's attractiveness depends on brightness associated with the encoded objective function.

The light intensity at a certain distancerfrom the light source obeys the inverse square law as given in Equation (14):

Furthermore,as air absorbs light,the light intensity becomes feeble as the distancerincreases.These two combined factors make most fireflies constraint to a limited distance,usually,several hundred meters at night,which is generally good enough for fireflies to communicate.The attractiveness of fireflyμat light absorption factorλis denoted byμλ:

whereris the distance betweenkthfirefly andlthfirefly andλis the light absorption factor andμ0is the initial attraction at distancerkl=0.The distancerklbetween two fireflies k and l at positionxkandxlis computed as given in the following equation:

wherexl,kis thekthcomponent of the positionxloflthfirefly.The fireflykattracted to another more attractive (brighter)fireflylis determined byyk:

In Equation (15),xkandxlare the positions of thekthandlthfirefly,andηis a randomization parameter.If the light absorption factorλ→0,thenμλ=μ0.Hence,flashing fireflies can be seen within the search domain.On the other hand,ifλ=∞,the attractiveness is almost zero,and the firefly moves randomly.When fireflies move randomly,there is plenty of chance for the FA to limit local maxima,resulting in premature convergence.As FA has these limitations,the randomness in FA replaced by GWO in the FAGWO algorithm.

The pseudo-code of the proposed FAGWO-based on optimum weight selection for feature weighing in the CBIR framework is shown in Algorithm 1.Here,the population of fireflies and grey wolves indicates randomly generated initial population of size m×d where m=0,1,2…M-1 and d is 1,2,…,z-1.

Algorithm 1 Pseudocode for FAGWO algorithm

4|CBIR USING GHMs AND FAGWOALGORITHM

In a primitive CBIR framework,the content is retrieved from the dataset in terms of features.The feature of a queried image is compared with the database's images using similarity measurement based on Euclidean distance.However,machine learning-based approaches are less cumbersome and with effective optimization technique,the highest is efficiency obtained in this CBIR system.

The basic steps involved in the proposed CBIR-FAGWO system described in Figure 3 illustrated as follows:

· Convert images in the database from RGB to monochrome images.

· Preprocess the query and images in the database with GLPF.

· Collect the GHM featuresf1,f2…fnof order (p+q) for each monochrome image to create a feature database.

· Compress features using PCA.Reduce the number of attributes fromNtoz.

· Multiply features with corresponding weights.The new feature matrix iswhere d is in range of 0,1,…,z-1.

· Feature database is segregated into training and testing set.

FIGURE 3 Proposed content-based image retrieval framework using FAGWO.FAGWO,firefly and grey wolf optimization;GLPF,Gaussian low-pass filtering;KNN,K-nearest neighbors;PCA,principal component analysis;RF,random forest

TABLE 1 Confusion matrix of two classes

TABLE 2 Precision,recall,and accuracy of classifiers KNN,RF with and without FAGWO

TABLE 3 Preset parameters of FAGWO

· Supervised learning classifiers,namely,RF and KNN are used.The training data is labeled and the input and output mapping is done,which is later used to classify the class labels for unseen instances accurately.Thus,the proposed algorithm learns about correct classes and the algorithm iteratively makes predictions on the training data.The Learning of the algorithm stops when the algorithm achieves minimum error.

· The confusion matrix is drawn between the actual and predicted classes.The accuracy is measured from the confusion matrix using the sum of the matrix's diagonal elements divided by 100.The fitness function of the FAGWO is the accuracy of the classifier.

· If the number of fireflies are’m,a random matrix array of sizem×dis created.Each string with a size ofdrepresents a possible solution to the optimization problem addressed.Associated with each string is a fitness value ofIcomputed by the evaluation unit.In the case of FA,the light intensity is the fitness function.A fitness value is a measure of the goodness of the solution that it represents.The operators in the optimization aim to transform the set of strings into strings with fitness values.Out of them,the fittestindividuals are selected for the next generation.Each random solution (weights of sized) is multiplied with the features to produce a weighted feature matrix.The random solution yields the highest fitness value and is retained in the next generation.The light intensity of each solution is computed and theλis initialized with a value between 0 and 2.If the condition isIk＜Il,thekthfirefly is attracted towards thelthfirefly and FA is executed.If the conditionIk＜Ilis satisfied,it is randomness in FA.At this condition,FA is replaced by GWO parametersa,A,andCare initialized.The weighted feature matrix is applied to the algorithm one after the other to find accuracy.The weights that produce the highest accuracy are retained for the next generation.For one generation,m fitness values are obtained.Out ofmfitness values,the three best fitness values are selected for the next generation in GWO asXα,Xβ,andXγ.The optimization process is terminated when the maximum generations ofgenerationscountis reached.The weights obtained at the maximum generation correspond to the optimized weights.When these optimized weights are applied,accuracies are maximized.In hunting,the global maxima is represented by catching the prey using Equations (11-13) in the hunting of grey wolves.

TABLE 4 RF classifier’s confusion matrix with FAGWO where 1—Africans,2—Beaches,3—Buildings 4—Buses,5—Dinosaurs,6—Elephants,7—Flowers,8—Horses,9—Mountains,10—Buildings

The accuracy obtained from the confusion matrix in the tabular form is given in Table 1.The parameters are true positive(TP),false negative(FN),false positive(FP),and true negative (TN).Retrieval efficiency (accuracy),precision,and recall can be obtained from the confusion matrix in the form of following equations.

FIGURE 4 Retrieved images from the database when query given is ‘African’

FIGURE 5 Retrieved images from the database when query given is ‘Dinosaur’

5|EXPERIMENTAL RESULTS

The proposed FAGWO is evaluated using The Corel database containing different pictures of indoor,outdoor,natural scenes and human-made objects.The images in the Corel-1K database are used for classification as shown in Figure 1.Because the pictures are in the RGB color standard,GHMs cannot compute on color images.So first,they are converted into monochrome images.Later the monochrome images are preprocessed with GLPF to obtain salient content of image information.For the order ofn,we obtainn2features.Hence,81 features of GHM of order nine have been used to create a feature database for each image.The NUMPY library in the python environment is used for the floating-point computation of GHMs polynomials.The features of GHMs are computed using the OpenCV library and feature dimensionality reduction is performed with PCA in the Scikit image.The compressed features are saved in a .CSV file.Scikitlearn models in Python are used for machine learning implementation.The number of reduced featureszis 10 in our case.The CSV file contains feature vectors imported to the scikit-learn package for training with a 70:30 train-test split.After the training,proposed weighed features are extracted for all the test images and fed to the KNN and RF models for validation and testing.

The code for swarm optimization algorithms,namely,FA and GWO is taken from Swarm Package Py 1.0.0.a5.The codes of these two optimization algorithms are rewritten to suit the requirements.For the obtained outcome (predicted class),a confusion matrix is drawn between the test images and predicted images to obtain accuracy,precision,and Recall.The FAGWO optimization technique ia used to increase retrieval efficiency.Random weights are initialized using the initial population of the FAGWO.The initial population of either fireflies or wolves is taken as 100 as depicted in Table 2.The best agents are selected and all the other agents change their position vectors using Equations(12-14).The FAGWO model runs for 20 generations and at the end of 20thgeneration,best precision,Recall,and retrieval efficiency are obtained for RF and KNN.The average precision,Recall,and retrieval efficiency results obtained are tabulated in Table 3.A percentage increase of 57.33% is observed in KNN while an increase of 9.2% is observed in RF classifiers retrieval efficiency when FAGWO is used along with GHM.

TABLE 5 Comparison of the precision of various state-of-the-art methods for different classes where 1—Africans,2—Beaches,3—Buildings,4—Buses,5—Dinosaurs,6—Elephants,7—Flowers,8—Horses,9—Mountains,10—Buildings

TABLE 6 Comparison of various state-of-the-art methods in terms of features used,fitness function and the computational time

Table 4 depicts the confusion matrix of RF classifier with FAGWO.KNN (without FAGWO) required better feature selection for obtaining separable classes.Hence,utilizing PCA and FAGWO are utilized for efficient weights and ultimately feature selection drastically increased the classifier's performance.The independent parameters were carefully tuned to obtain optimum performance in terms of global maxima.The retrieved images of the African class and Dinosaur class are shown in Figures 4 and 5.The proposed method is compared with state of art methods[15,21-23]in Table 5 and the results prove that the precision of the proposed method is very high.Our experimental results show that the retrieved image by the proposed technique has great accuracy against Corel database in comparison with[13]and[15].Table 6 shows a comparison in terms of features used,Fitness function and computational time.The computational time of [13,15] and proposed algorithm are 249,275.65 and 240.22 s for RF+FAGWO and 256.45 s for KNN+FAGWO.The computational time of RF+FAGWO is less compared with the existing methods[13,15].These results confirm the superiority of the proposed CBIR framework.

6|CONCLUSION

A CBIR system using GHMs with RF and KNN classifiers and a hybrid FAGWO algorithm is developed in this work.The fitness function using the FAGWO algorithm has accuracy from the confusion matrix.The optimized weights obtained from FAGWO significantly improved precision and recall.The exploration and exploitation are well balanced in the proposed hybrid FAGWO algorithms.In the future,the proposed method will be enhanced with deep convolutional neural networks and CBIR techniques will be enhanced to be suitable in applications such as fingerprint identification,biodiversity information,digital libraries,crime prevention,medicine,and historical research.

ORCID

Yasasvy Tadepallihttps://orcid.org/0000-0002-7684-5427