MODIFIED PARTICLE SWARM OPTIMIZATION BASED FEATURE EXTRACTION AND CLASSIFICATION FOR NEWSREADERS IMAGE-Deepa.V, Dr.Sengalliappan.M

MODIFIED PARTICLE SWARM OPTIMIZATION BASED FEATURE EXTRACTION AND CLASSIFICATION FOR NEWSREADERS IMAGE

Deepa.V1
Department of Computer Science
Research Scholar
Kovai Kalaimagal College of Arts & Science

Dr.Sengalliappan.M2
Research Supervisor
Dean of Computer Science
Department of Computer Science
Kovai Kalaimagal College of Arts & Science

Abstract:

Digital image are usually amid noise throughout the acquisition. Image denoising aims at reducing the noise without disturbing the image quality. The subject has been intensively researched through the past years and various algorithms were implemented with sensible results in noise reduction. This paper explains the processing of varied methods like image pre-processing, segmentation and have extraction model for any picture extracted from a video as frame.We propose a narrative approach supported patch matching using Gaussian mixture model (PDGMM) for image denoising and a neighborhood based segmentation supported horizontal vertical segmentation model for image segmentation. A unique technique for feature extraction supporting Particle swarm optimization (PSO), modified Particle swarm optimization and Speed-Up Robust Features (SURF). Our proposed algorithm provides promising result on extracting an image without loss in quality.

Keywords: Image Denoising; Patch Matching; Gaussian Mixture Model; Particle Swarm Optimization; Speed-Up Robust Features (SURF).

1. INTRODUCTION:

Digital image process lays emphasis on two main tasks: Betterment of pictorial information for the requirements of human interpretation and process of data image for storage, illustration, and transmission for freelance machine perception. The scope of digital image processing is to supply exact image without the loss in quality even after dimension changes.

Pre-processing involves the operations that are typically needed before to the analysis of data, extraction and are broadly classified as geometric or radiometric corrections. Radiometric corrections incorporate correcting the information for sensing element irregularities and undesirable sensing element or region noise, additionally as changing the information specified they exactly represent the emitted or mirrored radiation measured by the sensing element.

Image improvement involves modification of one or more elements of the image to increase clarity and details through the visual analysis and interpretation. It looks to switch the visual impact in a method that enhances the data content of the image. Digital image improvement strategies or techniques offer a bunch of selections for enabling the visual quality additionally for attractiveness of images [3] .The non-commissioned techniques supply a large range of means for altering image so on achieve images that are deemed visually adequate.

1. PREPROCESSING:

Image restoration is the process of recovering the underlying original image from its degraded form. The image captured through common imaging systems typically contain disturbance like noise, optical or motion blur attributable to sensing element limitations or other environmental conditions [8]. Due to the above mentioned limitations image restoration algorithms were used in applications like medical imaging, satellite imaging, surveillance, and general client imaging applications. Priors on natural image play an important role in image restoration. Image priors are familiar with denoise or regularize ill-posed restoration problems like deblurring and super-resolution.

III. FEATURE EXTRACTION:

3.1 PARTICLE SWARM OPTIMIZATION

The Particle Swarm Optimization (PSO) algorithm is proposed to model social behaviour of bird flocking or fish schooling. The algorithm is suitable for solving minimization or maximization problems [8]. Each element of the population is referred to as a particle, and is represented by three vectors and two real values: One of them points out the position of the particle, and the other one represents the last position change, and it is also called velocity vector. The third

one is a copy of the best position of the particle found as yet. Each particle moves in the solution space according to its current speed, the best solution found yet, and the position of the best of the population [2]. The PSO algorithm maintains several candidate solutions in the search space iteratively. In each iteration of the algorithm, each candidate solution is obtained by optimizing the objective function, determining the fitness of that solution. It has good performance even under post processing and pre-processing attacks (such as blurring, noise addition, rotation, JPEG compression).

3.2 SPEED-UP ROBUST FEATURES (SURF)

The algorithm can be described as a key point detector and descriptor. SURF approximates second order Gaussian derivatives with box filters [1]. With using integral images, image convolutions with these box filters can be calculated rapidly. Through the computed integral image, it is easy to calculate the sum of the intensities of pixels. The location and scale of interest points are selected by counting on the determinant of the Hessian matrix. Finally, the calculated maxima of the determinant of the Hessian matrix are interpolated in scale and image space. SURF constructs a circular region around the detected interest points via assign a unique orientation to gain invariance to image rotations. The orientation is computed using Haar wavelet responses in both x and y direction [9]. The Haar wavelets can be easily computed via integral images, similar to the Gaussian second order approximated box filters. After assigned the orientation, around the interest points the SURF descriptors are constructed by obtaining square regions. The windows are divided into 4 by 4 sub regions for keep in some spatial information. In each sub region, Haar wavelets are computed at regularly spaced sample points.

1. PROPOSED METHOD:

The input parameters are the test image and Hessian threshold. The output is number of matched key points, which isused as evaluation criterion to make detect decision. The higher criterion value means high performance. The Hessianthreshold parameter is determined with PSO [5]. The Particle Swarm Optimization (PSO) algorithm is proposed to model social behavior of bird flocking or fish schooling. The algorithm is suitable for solving minimization or maximization problems. Each element of the population is referred to as a particle and is represented by three

vectors and two real values: One of them points out the position of the particle, and the other one represents the last position change, and it is also called velocity vector. The third one is a copy of the best position of the particle found as yet. Each particle moves in the solution space according to its current speed, the best solution found yet, and the position of the best of the population [15]. The PSO algorithm maintains several candidate solutions in the search space iteratively and in each iteration of the algorithm corresponding to match for the current key point. We set t value to 0.6. RANSAC surveys the quality of the hypothetical model on input data set that is contaminated by outliers. For the analysis of copy-move forgery detection, the right matching should robust to the rotation and scaling, which is defined as location of point. So, for every outlier that are not appropriate for the translation matrix are removed with iteratively.

1. CLASSIFICATION:

5.1 MODIFIED BACK PROPAGATION ALGORITHM

To overcome some of the disadvantages of the Back propagation algorithm, two modification techniques are proposed in this paper. The first modification technique helps in reducing the mean square error and also reducing the number of epochs required for training the Artificial Neural Networks, hence speeding up the training of the Neural Networks [13]. The second modification technique computes the near optimum value of the learning rate for each training pattern that also reduces the number of training epochs. In the following, the two modification techniques are presented.

First Modification Technique: Second order approximation of the steepest-descent.

The back propagation method depends on the first-order approximation of the steepest descent method [12]. In the first modification technique the second-order approximation of the -descent method is proposed to beused instead of the first & approximation.

Second Modification Technique: Computation of Learning Rate per Pattern Using One Dimensional Optimization

This technique tries to get use near optimum learning rate within the specified range for each training input, grid search method is used, and this method requires only the objective function values but not the partial derivatives of the function (non-gradient method).

Third Modification Technique: Rounding Technique.

 This technique is used only for Neural Networks that are used in the classification, and triestotrain the Networks with virtually zero mean square error, this is done in the test phase and not in the training phase [18].

Modified Back Propagation Algorithm:

Step 0: Initialize weights (set to small random values).

Step 1: While stopping condition is false. Do steps 2 – 9.

Step 2: For each training pair, do steps 3 -9.

Step 3: Each input unit (, i = 1, …..) receives input signal and broadcasts this signal to all units in the layer above (the hidden units).

Step 4: Each hi- unit sums its weighted input signals,

Apply this activation function to compute the output signal,

Send this signal to all units in the layer above (output units).

Step 5: Each output unit sums its weighted input signals,

Apply the activation function to compute the output signal.

Step 6: Each output unit receives a target pattern corresponding to the input training pattern. Compute the error term.

Step 7: Calculate the near optimal learning rate using the second modification technique, which is mentioned above.

Step 8: Calculate the weight correction term and its bias correction. Send error term to units in the layer below.

Step 9: Each hidden unit sum its delta input (from the unit in the layer above). Multiply by the derivative of its activation function to calculate its error term Calculate the weight correction term, and calculate the bias correction term

(t)

Step 10: Each output unit updates its bias and weights, and each hidden unit updates its bias and weights.

Step 11: Enter the same input unit (, i = 1, …..) And repeat again steps from step 4 to step 10 except step 7

Step 12: (used in the test phase for classification only)

Step 13: Test stopping condition.

The Back Propagation algorithm is terminated at the weight vector W- when

Where is sufficient small error threshold.

The advantages of The Modified Back Propagation Algorithm:

• Minimizing the training time and the number of epochs required for training
• Searching for the optimum learning rate for each input pattern.
• Give virtually zero mean square error in thecase of classification problem.

1. RESULT:

Modified optimization based future extraction algorithm applying the image get the best future point from the images.

Table 1 show the parameters analysis of feature extraction for existing algorithm and proposed algorithm

This is below figure shows comparison of existing algorithm future extraction point (MPSO_HOG, SURF) is better prediction for proposed algorithm MPSO_SURF.

Feature extraction based Classification

Table 2 show the parameters analysis (accuracy, precision, recall measure, time period) existing algorithm and proposed algorithm

this figure show comparsion of accuracy,precision ,recall for existing (BPNN, HOGBPNN) algorithm and   proposed algorithm SURFBPNN

this figure show comparsion of time period for existing (BPNN, HOGBPNN) algorithm and   proposed algorithm SURFBPNN.

Table 3 show the parameters analysis (accuracy, precision, recall measure, time period) existing algorithm and proposed algorithm

this figure show comparsion of accuracy,precision ,recall for existing (BPNN,HOGMBPNN) algorithm and   proposed algorithm SURFMBPNN.

this figure show comparsion of time period for existing (BPNN, HOGMBPNN) algorithm and   proposed algorithm SURFMBPNN.

VII. CONCLUSION

This paper explains the processing of varied methods like image pre-processing, segmentation and have extraction model for any picture extracted from a video as frame. We proposed a narrative approach supported patch matching using Gaussian mixture model (PDGMM) for image denoising and a neighborhood based segmentation supported horizontal vertical segmentation model for image segmentation. A unique technique for feature extraction supporting Particle swarm optimization (PSO) modified Particle swarm optimization and Speed-Up Robust Features (SURF). Our proposed algorithm provides promising result on extracting an image without loss in quality.

VIII.REFERENCE

[1] Dong Hui, Han Dian Yuan,” Research of ImageMatching Algorithm Based on SURF Features ”,in International Conference on Computer Science andInformation Processing (CSIP),2012.

[2] Alec Banks, et al. “A review of particle swarmoptimization. Part II: hybridization, combinatorial, multicriteria and constrained optimization, and indicativeapplications”, Nat Compute (2008) 7:109-124, DOI10.1007/s11047-007-9050-z.

[3] Gheorghita Ghinea, (member, IEEE), Rajkumarkannan,and Suresh kannaiyan,”Gradient-Orientation-Based PCASubspace for Novel Face Recognition” in IEEE Access, vol2, 2169-3536, 2014.

[4] Hossein Sahoolizadeh, ZarghamHeidari, and HamidDehghani,” A New Face Recognition Method using PCA,LDA and Neural Network “in International Journal ofElectrical and Electronics Engineering 2:8, 2008

[5] Ajit P. Gosavi , S.R.Khot,” Facial ExpressionRecognition using Principal Component Analysis withSingular Value Decomposition” in International Journal ofAdvance Research in Computer Science and ManagementStudies, Volume 1, Issue 6, November 2013

[6] A. J. Fridrich, B. D. Soukal, A. J. Lukas, “Detection of copy-moveforgery in digital images”, Proceedings of Digital ForensicResearch Workshop, Trumansburg, 2003.

[7] Q. Wu, S. Wang and X. Zhang, “Log-Polar Based Scheme forRevealing Duplicated Regions in Digital Images,” in IEEE SignalProcessing Letters. vol. 18, no. 10, 2011, pp. 559-562.

[8] S.-J. Ryu, M. Kirchner, M.-J. Lee, H.-K. Lee, Rotation invariantlocalization of duplicated image regions based on Zernike moments, Trans. Inform. Forensics Secur. (T.IFS), 8 (8) (2013), pp.1355–1370

[9] H. Zhou, Y. Shen, X. Zhu, B. Liu, Z. Fu, and N. Fan, “Digitalimage modification detection using color information and itshistograms,” Forensic Science International, vol. 266, pp. 379–388, 2016.

[10] S. Wenchang, Z. Fei, Q. Bo and L. Bin, “Improving image copy-move forgery detection with particle swarm optimizationtechniques,” in Proc. in China Communications, vol. 13, no. 1, pp.139-149, Jan. 2016.

[11] F. Yang, J. Li, W. Lu, J. Weng, ‘Copy-move forgery detection based on hybrid features’, Engineering Applications of ArtificialIntelligence, Volume 59, 73-83.

[12] Q. Wu, S. Wang and X. Zhang, “Log-Polar Based Scheme forRevealing Duplicated Regions in Digital Images,” in IEEE SignalProcessing Letters. vol. 18, no. 10, 2011, pp. 559-562.

[13] V.V.Joseph, Rajapandian, N.Gunaseeli Modified Standard Backpropagation Algorithm with Optimum Initialization for Feedforward Neural Networks” JISE, GA, USA, ISSN: 1934- 9955, vol.1, no.3

[14] M. Emam, Q. Han, and X. Niu, “PCET based copy-move forgery detection in images under geometric transforms,” Multimedia Tools andApplications, vol. 75, no. 18, pp. 11513–11527, 2016.

[15] V. Christlein, C. Riess, J. Jordan, C. Riess, and E. Angelopoulos, “An evaluation of popular copy-move forgery detection approaches,” IEEE Trans.Inf. Forensics Security, vol. 7. Dec. 2012.

[16] Y. Zhu, X. Shen, H. Chen, Copy-move forgery detection based on scaled ORB, Multimed Tools Appl, pp. 1-15, DOI: 10.1007/s11042-014-2431-2.

[17] I. Amerini, L. Ballan, R. Caldelli, A. Del Bimbo, L. Del Tongo, G.Serra (2013) Copy-move forgery detection and localization by means of robustclustering with J-Linkage. Signal Process Image Commun 28(6):659– 669.

[18] Omaima N.A. AL-Allaf, “Improving the Performance of Backpropagation Neural Network Algorithm for Image Compression/Decompression System”, Journal of Computer Science 6, 2010.