An Efficient Gabor Walsh-Hadamard Transform Based Approach for Retrieving Brain Tumor Images From MRI

Brain tumors are a serious and death-defying disease for human life. Discovering an appropriate brain tumor image from a magnetic resonance imaging (MRI) archive is a challenging job for the radiologist. Most search engines retrieve images on the basis of traditional text-based approaches. The main challenge in the MRI image analysis is that low-level visual information captured by the MRI machine and the high-level information identified by the assessor. This semantic gap is addressed in this study by designing a new feature extraction technique. In this paper, we introduce Content-Based Medical Image retrieval (CBMIR) system for retrieval of brain tumor images from the large data. Firstly, we remove noise from MRI images employing several filtering techniques. Afterward, we design a feature extraction scheme combining Gabor filtering technique (which is mainly focused on specific frequency content at the image region) and Walsh-Hadamard transform (WHT) (conquer technique for easy configuration of image) for discovering representative features from MRI images. After that, for retrieving the accurate and reliable image, we employ Fuzzy C-Means clustering Minkowski distance metric that can evaluate the similarity between the query image and database images. The proposed methodology design was tested on a publicly available brain tumor MRI image database. The experimental results demonstrate that our proposed approach outperforms most of the existing techniques like Gabor, wavelet, and Hough transform in detecting brain tumors and also take less time. The proposed approach will be beneficial for radiologists and also for technologists to build an automatic decision support system that will produce reproducible and objective results with high accuracy.


I. INTRODUCTION
Brain tumors are one of the biggest killers of children and adults under 40 worldwide.A brain tumor is a cluster of abnormal cells that grow out of control in the brain.People with a brain tumor experience a variety of physical symptoms including headaches, vision problems, seizures, memory loss, changes in personality, difficulty in concentration, loss of coordination and changes in speech, loss of balance and The associate editor coordinating the review of this manuscript and approving it for publication was Amin Zehtabian .mood swings.There are two types of brain tumors: cancerous (malignant) and noncancerous (benign).Malignant brain tumors are cancers that typically grow faster than benign tumors, and quickly invade surrounding tissue.When benign or malignant tumors grow, they can cause the pressure inside the skull to increase.This can be the reason of brain damage, which is life-threatening.Brain tumors substantially affectquality of life and change everything for a patient and their loved ones.The American Cancer Society's estimated for 2021 that about 18,600 people (10,500 males and 8,100 females) will die from brain and spinal cord tumors in FIGURE 1. MRI study scans [2].
the USA [1].Magnetic resonance imaging (MRI) is the most popular technique for diagnosing brain tumor.MRI presents tremendous detail of brain, spinal cord, and vascular anatomy.These three visualized anatomy will have all three orientations: axial, sagittal, and coronal.The most popular MRI sequences are T1-weighted, T2-weighted scans, and Flair (Fluid Attenuated Inversion Recovery).The T1 weighted brain image also can be taken in 3D visualization of anatomy in all three planes: Axial, coronal and sagittal as in Fig. 1.The axial orientation of the MRI head image is considered from neck to head.The coronal orientation starts on the tip of the nose and ends behind the head.The sagittal orientation extends from ear to ear.
Medical imaging modalities which include computed tomography (CT), magnetic resonance imaging (MRI), ultrasound, and positron emission tomography are nondestructive [2].They consequently help in capturing the images of internal parts of organs [3], [4].In the Content Based Image Retrieval (CBIR) domain for brain image, there are two major concepts in research works.One kind of method focuses on automatic retrieving images from Picture Archive and Communication System (PACS) like databases, which search images of the same imaging modality, body orientation, body region [5].Another kind of method put their efforts into retrieving images that characterize the similar disease convenient for diagnostic comparing [6]- [8].Manual MRI retrieval from a massive archive of imaging data with similar structures is a difficult and challenging task for radiologists.It relies upon the availability and understanding of the radiologist, who examines MR images and retrieves the relevant images from the archived data.This manual retrieval technique is impractical, non-reproducible, and time-consuming for a massive quantity of archived data.
To address this problem, automated CBIR is a probable solution for indexing archived images with minimal intervention through radiologists.In this research, we focus on CBIR for the retrieval of brain tumors images.Especially, while the radiologist provides a query image, the CBIR system retrieves the same pathological kind of brain tumor images from the database, then, the radiologist selects the maximum closely associated retrieved images and useful for diagnosis and treatment of the current case [9].There are more than 120 forms of brain tumors that have been diagnosed by the World Health Organization (WHO) consisting of astrocytoma, gliomas, meningioma, medulloblastoma, pituitary tumors, and so on.These tumors are classified based on their location, grade, and nature of tumor cells.The clinical community specializes in glioma because the adults are focused on the same.Treatment of glioma relies upon its location, shape, and size [10].
In this research, we have designed a content-based system for retrieving Glioma brain tumor from MR images.Contentbased image retrieval system uses the information presented in input image data and creates visual attributes of its content.In this proposed system we involve several filtering methods like Mean Filter, Median Filter, Conservative Filter and Crimmins Speckle Removal for removing noise and we adopt Gabor Walsh-Hadamard Transform strategy for feature extraction.Finally, we use Fuzzy C means with Minkowski to measure similarity distance for efficiently identify brain tumor using the obtained features.The proposed method is evaluated on a publicly available CE-MRI dataset, whole brain atlas dataset, IBSR web service and so on.We performed extensive experimental works for glioma brain tumor image retrieval from MRI and also test robustness, evaluated the performance of the proposed method, and compared our results with state-of-the-art brain tumor retrieval on the same dataset.The main contributions in this work are: 1. To retrieve the exact matching image from the dataset, implement noise removal using Mean Filter, Median Filter, Conservative Filter, and Scimmins Speckle Removal.
2. We adopt Hybrid of Gabor and Walsh-Hadamard Transform or feature extraction.
3. The proposed work is implemented on CE-MRI dataset, and IBSR dataset for content-based brain tumor retrieval using Fuzzy Clustering with Minkowski Distance.
The rest of the paper is organized as follows.Section 2 discusses the related work, Section 3 discusses proposed research framework and methodology in detail.Section 4 contains feature matching process.The experimental settings, retrieval performance, results and comparisons are shown in Section 5.And the conclusion is presented in Section 6.

II. RELATED WORK
The application of computer vision technique is Content-Based Image Retrieval (CBIR) system which retrieves images from large databases.CBIR is also known as Query by Image Content (QBIC).Retrieval of similar images from the large datasets based on extracting some features of color, texture and shape and enable indexing the images automatically [11].Content-based image retrieval [12] has proposed a region of interest (ROI) for retrieving strategy.ROI is used as query input, then retrieves the tumor images from large dataset by using SIFT features.Local Binary Patterns (LBPs) were successfully used in Image CLEF med, 2D Hela, for texture feature extraction of brain MRI and it is retrieval tasks are implemented [13]- [15].
Content-based brain tumor retrieval on the T1-weighted contrast-enhanced MRI (CE-MRI) dataset [22]- [24] proposed a margin information descriptor (MID) as a feature extractor.They use tumor region for querying the database to retrieval the same.Apart from traditional features such as shape, intensity etc., in this research they use margin information descriptor (MID).MID helps to define the tissues surrounding the tumor for describing the image content.Maximum mean average precision projection is used to find approximate value to improve retrieval performance.
By using region-based technique in the extraction of shape feature from Computer Tomography (CT) images are done using CBIR concept.Similarly, for classifying and retrieving the different medical images in CBIR using Support Vector machine (SVM) technique with different distance metric measures were used for analysis the result in terms of recall and precision [25].In Deep Belief Convolutional Neural Network (DBCNN) concept CBIR system is used for retrieval of medical images.For pre-processing median filter is used [26].
In the pre-processing stage, Gradient vector field model and Median filter techniques are used for boundary detection.For designing CBIR in medical images Steerable filter is used for extracting feature of texture and for representing features of texture by using PCA technique.For retrieving the similar images from the large data set using Euclidean distance is described in article [27].The segmentation of tumor tissues in the MRI Brain image using FCM and K-means clustering is performed.In the performance analysis K-means clustering performed better result [28].To segment the Glioma tumor image, for the training process cascaded CNN architecture was implemented using patches [29].To identifying the brain tumor Convolution encoder network (CEN), CNN, long short-term memory (LSTM), dual-force CNN and U-Net CNN were implemented [30]- [34].
Few surveys of important existing techniques are described in Table .1.
The cells in abnormal condition react to form uncontrolled cell division in the brain is called cerebral tumor [39].It replicates and destroys healthy cells in the human body.Grade 1 and grade 2 are low grade tumors.Grade 3 and grade 4 are high grade tumors.In this research, they classify normal and abnormal tumor brain from the MRI database.They use various wavelet transforms and support vector machines for accurate classification.The CNN with multimodal information fusion [40] is used to obtain lesions of the brain under 3-D space.Compared to 2-D techniques, this method uses less fault input for processing.Though it is advanced, computational complexity in detection is high in this technique.
The multi-level features extraction [41] and concatenation for detecting brain tumor at the earliest is discussed.Two deep learning methods such as Inception-v3 and DensNet201 are used in efficient feature extraction process.Finally, outputs are concatenated and passed to the softmax layer for accurate classification.The main problem is more tough technology is implemented and the cost is high.Genetic algorithm [42] is used to detect an optimized edge in medical dataset.The algorithm called optimal threshold value is used for finite edge detection based on minimized cost.Feature of an input image is improved by using Balance Contrast Enhancement Technique (BCET) for providing better medical images with fine characters.In this, BCET cannot provide all features in superior quality.This technique is only useful for edge detection strategy.Edge detection cannot provide an accurate classification of all brain tumors.

III. PROPOSED METHODOLOGY
This study introduces a content-based image retrieval system for automatically identifying brain tumor from a large MRI images dataset.In this proposed work, we design a new framework for texture feature extraction by using a hybrid of Gabor Walsh-Hadamard Transform (GWHT).To retrieve similar brain tumor images for the input query image from the database, the distance metric measure is used for matching.To measure the distance for the similarity between query image and database images, we propose a distance metric of Fuzzy Clustering with Minkowski distance that has been newly introduced in this study.The optimization of the prediction system is designed into three phases such as.
Phase 1 (Pre-Processing): We remove noise from raw MRI dataset involving Mean Filter, Median Filter, Conservative Filter, and Scimmins Speckle Removal.Phase 3 (Retrieval of Brain Glioma Tumor Image): We retrieve brain tumor images from MRI database using Fuzzy Clustering with Minkowski Distance.The architecture of the proposed system is shown in Fig. 2.

A. PRE-PROCESSING OF MRI IMAGES
The aim of pre-processing is to improve the high quality of the MRI brain images and make it in a form suited for further processing.In this work, we implemented image denoising to take away noise from brain image, which will restore the actual image [43].For that purpose, we used Mean filter, Median filter, Conservative Filter, Crimmins Speckle Removal for producing effective results.Brain tumor images are very tough to preprocess using a single filter.Single filter output has more error features which makes the wrong prediction is existing research solution.Our multiple preprocessing step filters the noises accurately for further effectiveness.Each filter is described in detail in the following subsections.

1) MEAN FILTER
The mean filter is used to blur an image so that you can remove noise.It includes determining the mean of the pixel values in the form of a M × N kernel.The mean filter helps to reduce the variation of intensity between one pixel to next pixel.In this filter, the value of each pixel is replaced by the mean value calculated in a local neighborhood.Mean filter have the power to discard unrepresentative pixel values of their surroundings.The intensity value of the center pixel value is replaced by the mean value.Thus, it removes the noise in the image and we get smooth edges of the image.For a filter of length N, any filtered data point can be represented mathematically in terms of the average of the last N measured data points as The mean filter can also be treated as a convolution of the measured signal with a vector of N constant coefficients, each equal {1/N}.For applying this filter in Python, by using Open-CV library to get a mean filter of the image.The resultant of applying mean filter is given below in Fig. 3.

2) MEDIAN FILTER
The median filter, scan the entire image, using a 3 × 3 and recalculate the value of the center pixel by simply taking the median of all of the values inside the matrix, that is replaces it with the median of those values.Main advantage of using median filter is, it calculates robust average value than mean filter.Generally median value is any one of pixel value around neighbors.Therefore, it cannot create unrealistic value in pixel which help to preserve the sharp edges of images.This is very significant feature for brain tumor image.
where, Z represents a neighborhood pixel values defined by the user and the centered around location [p, q] in the image.The resultant of applying median filter is given below in Fig. 4.
Similarly, if it is less than the minimum value then it is by the minimum value in the output image.The parameters which are used in conservative filter is neighborhood size, or filter size, is specified in the m and n dimensions.These dimensions should be odd, positive integer values (for example: 3, 5, 7, 9, etc.).Fig. 5 shows the resultant of applying the conservative filter.

4) CRIMMINS SPECKLE REMOVAL
The Crimmins algorithm is used to remove speckle noise and smooth the edges.This algorithm compares the intensity of a pixel in an image with the intensities of its 8 neighbors.The algorithm considers 4 sets of neighbors (N-S, E-W, NW-SE, NE-SW).Let p, q, r be three consecutive pixels.Then the algorithm is: For each iteration do the following steps.a) Dark pixel adjustment: For each of the four directions 1) Process whole brain image with: if p ≥ q + 2 then q = q + 1 2) Process whole brain image with: if p > q and q ≤ r then q = q + 1 3) Process whole brain image with: if r > q and q ≤ p then q = q + 1 4) Process whole brain image with: if r ≥ q + 2 then q = q + 1 b) Light pixel adjustment: For each of the four directions 1) Process whole brain image with: if p ≤ q − 2 then q = q − 1 2) Process whole brain image with: if p < q and q ≥ r then q = q − 1 3) Process whole brain image with: if r < q and q ≥ p then q = q − 1 4) Process whole brain image with: if r ≤ q − 2 then q = q − 1 The resultant of Crimmins Speckle Removal is given in Fig. 6.

B. FEATURE EXTRACTION
The main issue in CBIR is to extract the features which include texture, shape, etc. of the brain image efficiently and then represent them in a particular form to be used effectively in the matching of images.In our proposed work, we implement texture feature extraction by using hybrid of Gabor Walsh-Hadamard Transform (GWHT) technique.Gabor is a multi-scale, multi resolution filter [44].In Gabor filter pixel value x and y denotes the position of the image.Here ω denotes center frequency, θ denotes the Gabor's orientation of direction and σ shows that standard deviation of Gaussian function with x, y axis of the image.
where, ω x0 , ω y0 -center frequency of x and y directions of the image.
σ x , σ y -standard deviation of the Gaussian function with x and y axis or direction.
x, y -Position of the image in pixel format.Replace the Equation (3) as × e j(ω x0 x cos θ k +ω y0 y sin θ k ) (5) ) In this work, applied the orientation θ of Gabor direction as 0 • , 20 • , 40 • , 60 • , 120 • with their frequency values of 60, 80, 120, 140.After applying the Gabor Transform to the brain image, the texture features are extracted, in this output applying the Walsh-Hadamard Transform (WHT) to get more accurate and efficient result.WHT is based on correlation between local pixels of the brain image.Walsh transform matrix can be defined as WT i , i = 0, 1, . . .N − 1.The propoerties of WT are given below: 1.The values of Walsh Transform matrix is (WT i ) +1 and −1.

WT i
The Walsh transform matrix's row is equal to the row of Hadamard matrix and it is defined by index value of Walsh which is range from 0 to N-1.The Walsh transform matrix's row is equal to the row of Hadamard matrix and it is defined by index value of Walsh which is range from 0 to N-1.The properties of Hadamard matrix are given below: To change the order of Hadamard matrices by permuting rows and columns and also multiplying by the value −1 in rows and columns.The 4 × 4 matrix is defined as Walsh Hadamard Transform (WHT) is defined by sparse factorization of Walsh transform matrix and each factor value is referred as stage.In the WHT the input and output value of each stage is defines as factor value of decomposition.The sparse factorization of identity matrix is obtained from HD matrix with its inverse function.
To extract the texture feature and gives more clarity to the values by sum of the selected kernel values with projected onto the blocks of the image.These projection of selected kernel values for the blocks of the image is computed as following: where b i is the i th block of the image.p 1 , p 2 , p 3 , p 4 are projection of selected kernels {k 0 , k 5 , k 10 , k 15 } of the blocks of the image.
The sum of all projected kernels of the blocks image is calculated as follows: is consider as pixels and construct Gaussian adjacency matrix A and Diagonal Matrix D. 11.Construct Laplacian Matrix LMusing Gaussian adjacency matrix A and diagonalmatrix D. 12.To create unit matrix k by using denormalise the matrix LM. 13.For each row of unit matrix K; using K-means clustering, create a clusters.In which the value of pixels which is closest to the cluster centroid value.14.Collect all pixel values in the K-clusters.15.This collected pixel values are consider as tissues of brain tumor.

C. RETRIEVAL OF BRAIN GLIOMA TUMOR IMAGE FROM MRI DATABASE
In this work we used Fuzzy C-Means (FCM) algorithms with Minkowski distance to retrieve the similarity Glioma tumor images from the large image dataset [45].Main objective of using this algorithm is, it minimizes the Euclidean distance between the data and cluster center.The FCM is stated by: where, U = µ i,j KXN ∈ [0, 1] where, U is partition matrix, A i is optimization variable used in local norm of matrix, X is set of non-labelled data, By using (22)(23)(24) Fuzzy C means with Minkowski distance.The most common method for comparing two images in content-based image retrieval is by using an image distance metric measure.For example, a distance of 0 signifies an exact match with the query.The retrieved images are ranked in ascending order based on their relevance.

IV. RESULT AND DISCUSSION
Dataset Description:For the experimental purpose 1500 MRI brain images are taken from MRI scan center and some images are from IBSR web service developed through CMA at Massachusetts General Hospital and the whole brain atlas [46].Some images are collected from CE-MRI dataset.Almost all images are collected from 350 patients with a total number of 4500 images, but in this work, we proposed T1-weighted images with a default size of 256 × 256.We trained all images in the dataset by applying CBIR steps of pre-processing, feature extraction and feature matching process.And also repeat the same processing steps for the query image.

A. ANALYSIS FOR PREPROCESSING WORK
For removing noise and for smoothening images, we implemented Mean Filter, Median Filter, Conservative Filter, and Crimmins Speckle Removal using Python and investigated their effectiveness.Mean Filter and Median Filter resultshaveshowed bad filter performance for MRI image quality.The results of the Conservative filter are better filter compared to the mean filter and median filter.But smoothening of edges of the tumor image is not detect properly.So Crimmins Speckle Removal result gives better performance in terms of removal of noise and smoothening of the edges.
The peak signal-to-noise ratio (PSNR) is an engineering term defined as the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation.Because many signals have a very wide dynamic range, PSNR is typically expressed in terms of the logarithmic decibel scale.It is most simply defined via the mean squared error (MSE) value for two m × n monochrome MRI brain images P and Q is defined as:  From Tab.2, it observed that our Crimmins Speckle Removal filter for removal of noise and smoothening of the image is better than other filtering techniques on the data set.

B. ANALYSIS RESULTS OF FEATURE EXTRACTION
This section discusses the experimental results of feature extraction carried out using Python.The results of the proposed GWHT feature extraction method are presented comparing the other existing three methods of Gabor Transform, Wavelet Transform, and Hough Transform are proposed in [47]- [50].We used five metrics for analyzing the efficiency of the proposed method to detect the tumorous brain image.There are false alarm (FA), missed alarm (MA), sensitivity, specificity and accuracy.The classification error rate measured in terms of spurious detection of tumor images (FA) and fails to identify the tumor images (MA) are given in (28) and (29).The accuracy of the classification is measured by using the sensitivity, specificity, and accuracy are given from (30) to (32).The observed results from the Tab. 3 prove that the proposed method outperforms better than other existing approaches.The proposed method obtains 98% of accuracy on predicting the tumor images from the dataset.The resultant image is given in Fig. 8.

C. ANALYSIS RESULTS OF RETRIEVAL OF BRAIN IMAGE
This section presents the results of retrieval of brain image where precision and recall are used to measure for retrieval performance evaluation measures [22], [24].

Precision =
No.of relevant_images retrieved Total_No.ofimages_retrieved (33) Using this fuzzy c-means with Minkowski distance metric measure in (22), we can retrieve very closed similarity Glioma tumor images from the dataset.
Distance of 0 signifies an exact match with the query, and it is organized in rank order.Fig. 11 shows the result of retrieval of brain image from large data set by using fuzzy c-means with Minkowski distance metric measures.By applying these measures in (33) and (34).Fig. 11 shows the comparative performance of retrieval of the brain image.
The above Fig. 9 shows that the retrieved images are ranked in ascending order based on their relevance to the given input.The top 5 most similar retrieved images which is the precision at the position where 5 is the most similar database images      in retrieval of brain tumor images from the large database.
The experimental results show that this method can improve the results with an accuracy of around 94%. Fig. 11 shows time taken by all techniques and comparative study is performed.Our proposed takes only 30ms to compute the feature and identify the tumor features in the input-image.
Hence, our proposed algorithm, hybrid of Gabor Walsh-Hadamard Transform (GWHT) feature extraction and retrieval of images from the data set by using Fuzzy C means with Minkowski distance metric measures outperforms better in terms of accuracy, execution time, and efficient retrieval of images.This GWHT is more reliable and consume less time for extracting features.The super pixels are used to recover the tumor in an effective way and it proves better algorithm than any other existing algorithm.This will be the best method to predict the tumor patient efficiently and helpful to the medical industry for the diagnosis and early prediction of Glioma tumor patients.

V. CONCLUSION
Content-based image retrieval is challenging among recent image processing researchers.In this research article, we propose a new CBIR approach based on GWHT feature extraction technique for brain tumor retrieval.The proposed method yields better performance for retrieving the medical image with 97.9% of accuracy.Result evaluation shows our proposed technique is superior in precision and recall compared to other existing techniques.In addition, the proposed method took less time compared to other feature extraction methods.Retrieving nearest image helps radiologist to predict true positive results as early as possible.The limitation of the work is we cannot segregate false positive images in case of high similarity among the image pixels.Retrieving the relevant image based on features is considered as a first and foremost essential step for medical diagnosis.When this technique works with high accuracy, further processing techniques can predict diseases as early as possible.In future, this limitation can be achieved by some semantic-based similarity calculation techniques.Also for another limitation like high similarity of pixels of the image cannot be segregate as false positive; therefore, it can be solved by using optimization algorithm in Artificial intelligence or swarm intelligence.

FIGURE 2 .
FIGURE 2. Proposed GWHT architecture.Phase 2 (Feature Extraction): We extract significant characteristics of brain tumor exploring several feature extraction methods such as Gabor Transform, Wavelet Transform, Hough Transform, Hybrid of Gabor and Walsh-Hadamard Transform.Phase 3 (Retrieval of Brain Glioma Tumor Image): We retrieve brain tumor images from MRI database using Fuzzy Clustering with Minkowski Distance.The architecture of the proposed system is shown in Fig.2.

FIGURE 3 .
FIGURE 3.An example image of Mean Filter.

FIGURE 4 .
FIGURE 4.An example image of Median Filter.3)CONSERVATIVE FILTER smoothening the image and also reducing the noise in an image.It works by calculating the minimum and maximum neighboring values surrounding a grid cell value.If the intensity value of center pixel value is greater than the calculated maximum value, it is replaced with the maximum value in the output image.Let 4 × 4 window size of image.Take mid value of 4 × 4.

FIGURE 5 .
FIGURE 5.An example image of Conservative Filter.
HDR n = Ra n HDR n −1 (12) where HDR n -Walsh Hadamard Transform with radix R Ra n -Factorization of radix Ra n -Number of input element The WHT consists of Fourier and Cosine Transforms in the basic functions that are a set of orthogonal sinusoidal waveforms.The WHT applied in Squared size gallery space images generating M × M blocks from each image for texture feature extraction [41].Here we are using 4 × 4 blocks.Instead of considering the whole pixels of the image selectively choose the super pixels by using clusteringmethod.These selected pixels are considered as kernel of M × M blocks.The texture features are extracted by projecting sum of selected kernels {k 0 , k 1 , k 2 , k 3 , • • • k 15 } of WHT on theblocks of the image.The diagonal kernels are {k 0 , k 5 , k 10 , k 15 } are selected for extracting the texture features of the image.

VOLUME 9 ,Algorithm 1
2021 where s = 0, 5, 10, 15 are the diagonal kernel values of the image.Gabor and Walsh-Hadamard Transform (GWHT) Input: MRI brain tumor Output: Texture Feature Extraction of tumor tissue of brain 1.Read the brain image as in binary form from the data set.2. Applying the two-dimensional Gabor filter for the image by using Equation (4).3. Select 16 × 16 image size of each block and applying the Fourier transform.Then select the Fourier transformed Gabor size of 8 × 8 of image having 5 orientations and 4 scale values.4. Apply the Walsh-Hadamard transform, Partition each image into equal sized blocksof 4 × 4. Im = n i=0 bl i where n denotes the number of blocks.5. Take the diagonal kernel values are {k 0 , k 5 , k 10 , k 15 } for extracting the texture feature of the image.6. Project the selected kernels of WHT on the blocks 4 × 4 of the image s = n i=1 bl i * k s .Where bl i represents i th block bl and k n represents n th kernel K and n denotes the number of blocks.7. Calculate the texture strength of the image by using the formula T = bl 2 n − k 2 .The pixel intensity value of n th block of bl image and the k represent the projected kernel values of n th block of image.8. Repeat the steps 6 & 7 for all blocks of the image calculate the texture strength of the image.9.Each Blocks texture strength is taken as the super pixel s 1 , s 2 , s 3 , • • • n of the image.10.By using super pixels s 1 , 2 , s 3 , • • • , s m of the image

FIGURE 8 .
FIGURE 8. Feature extraction of proposed method.
returned.Fig.10shows the comparitive performance measure using GWHT -Fuzzy C-Means with Minkowski distance methods and other existing techniques.The evaluation results from Fig.10, shows GWHTfuzzy C-means with Minkowski Distance as a retrieval method that has good accuracy, good precision of 94.12%,

FIGURE 10 .
FIGURE 10.Comparative analysis of existing vs proposed methodology.

FIGURE 11 .
FIGURE 11.Time taken for image retrieval.

TABLE 1 .
Survey of existing techniques.

TABLE 2 .
Average PSNR value for different types of filters.