Neural Network Ensembles For Image Identification Using - - PowerPoint PPT Presentation

Neural Network Ensembles For Image Identification Using Pareto-optimal Features

Monday, 7th July 2014

w.albukhanajer@surrey.ac.uk www.surrey.ac.uk/computing

Wissam A. Albukhanajer, Yaochu Jin and Johan A. Briffa 2014 IEEE World Congress on Computational Intelligence July 6-11 2014 Beijing, China IEEE Congress on Evolutionary Computation

Department of Computing Faculty of Engineering and Physical Sciences University of Surrey United Kingdom GU2 7XH

In this talk:

Introduction :

• What is RST invariant?
• Trace Transform - Invariant Feature Extraction

Evolutionary Trace Transform with Noise

• Multi objective optimisation
• Pareto front

An Ensemble Classifier

• An Ensemble with different features
• Majority voting

Experimental Study

• Robustness to Scale
• Robustness to RST and Gaussian noise
• Robustness to RST and Salt & Pepper noise

Summary and Conclusions

What is RST Invariant Feature?

The Rotation, Scale and Translation (RST) Invariants

250 750 1250 1750 500 1000 1500 130 150 170 120 140 160 180

A B

A B

Introduction

Inspired by Radon transform, proposed by Kadyrov and Petrou [1]

[1] A. Kadyrov and M. Petrou, "The Trace transform and its applications," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, pp. 811-828, 2001.

Trace Transform is a generalisation of Radon Transform. The functional calculated on the image pixels is not necessary the integral.

Introduction The Trace Transform (TT)

Construction of a Triple feature:

Figure: Trace Algorithm. A Triple feature from an Image.

The Trace Transform (TT) (Cont.)

Different Trace Transforms can be produced Using different trace functionals:

Figure: A Butterfly image (a) and its Trace transforms using different Trace functionals: (b) Gradient, (c) Integral and (d) Standard deviation.

Angle (ϴ) (p)

(d)

(p) Angle (ϴ)

(b) (a) (c)

Angle (ϴ) (p)

The Trace Transform (TT) (Cont.) Introduction

The Objectives are to minimize the following:

∑ ∑∑

= Ξ Ξ = = Ξ

− = − Ξ =

K k k b K k C j k jk w

S S

k

1 2 1 1 2

) ( ) ( µ µ µ

Multi-Objective Optimisation (MOO)

ε : is a small quantity to avoid division by zero.

where:

Evolutionary Trace Transform with Noise (ETTN)

: Mean of all classes of Ξ Triple feature : Mean of class with Triple feature Ξ : Number of samples in class k : Number of classes Ξ :The Triple feature of class

∑ ∑

= Ξ Ξ = Ξ

= Ξ =

K k k C j jk k k

K C

k

1 1

1 1 µ µ µ

Ξ k

µ

Table I NSGA-II Parameter Set-up Figure: Pareto front. Three non-dominated solutions are chosen from the optimal front

NSGA-II and Pareto Front

TABLE II: Triple Features Combinations From Evolutionary Trace Transform Algorithm (ETTN) in

• Fig. 2

TABLE III: Functionals Description TABLE IV: Pairs of Triple Features Combinations From ETTN Algorithm

Funtctionals Description

Ensemble of classifiers

Figure: Structure of the ensemble classifier using Pareto optimal feature set.

Majority voting: Similar MLP base classifiers with different features as inputs

[1] A. Kadyrov and M. Petrou, "The Trace transform and its applications," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, pp. 811-828, 2001.

Experimental Results

Figure: Fish database [1].

Each image subjected to random RST deformation + Noise

Robustness to Scale Only

Figure: Robustness to scale only (one test sample per class for each scale factor).

[1] A. Kadyrov and M. Petrou, "The Trace transform and its applications," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, pp. 811-828, 2001.

Figure: Robustness to Gaussian noise of zero mean and standard deviation 2 = 0, 2, 4, 6, 8 and 10, of each approach, (Fish-94 database). Performance are shown when the object is scaled from 1 to 0.3, rotated and translated in a random way, and Gaussian noise has been added to the whole image with standard deviation values corresponds to each figure (one test sample per class for each scale factor).

Robustness to Gaussian Noise

Robustness to Gaussian Noise

Robustness to Gaussian Noise

Robustness to Salt and Pepper Noise

Figure: Robustness to additive salt & pepper noise with percentage of altered pixels % = 0, 1, 2,...,6 of each approach, (Fish-94 database). Performance are shown when the object is scaled from 1 to 0.3, rotated and translated in a random way, and noise has been added to the whole image with noise levels corresponds to each figure (one test sample per class for each scale factor).

Robustness to Salt and Pepper Noise

Robustness to Salt and Pepper Noise

Summary and Conclusions

• We presented an ensemble classifier using a set of Pareto-optimal Trace

transform features.

• While the traditional Trace transform that uses thousands of features, the

single classifiers or ensemble classifiers using the features extracted by the evolutionary multi-objective Trace transform (ETTN) are able to accurately classify noisy RST deformed images with a much lower computational cost.

• Our results indicate that different Pareto-optimal features can introduce

diversity in the ensemble classifier. As a result, no particular effort is needed to generate diverse base classifiers.

Publications

• W. A. Albukhanajer, Y. Jin, J. Briffa and G. Williams “Evolutionary Multi-Objective

Optimization of Trace Transform for Invariant Feature Extraction” IEEE Congress on Evolutionary Computation CEC, Brisbane, Australia, June 10-15, 2012.

• W. A. Albukhanajer, Y. Jin, J. Briffa and G. Williams “A comparative study of multi-objective

evolutionary trace transform methods for robust feature extraction,” in Evolutionary Multi- Criterion Optimization, ser. Lecture Notes in Computer Science, R. Purshouse, P. Fleming,

• C. Fonseca, S. Greco, and J. Shaw, Eds. Springer Berlin Heidelberg, 2013, vol. 7811, pp.

573–586. [Online]. Available: http://dx.doi.org/10.1007/978-3-642-37140-0 43

• W. A. Albukhanajer, J. A. Briffa, and Y. Jin, “Evolutionary Multiobjective Feature Extraction in

the Presence of Noise,” Submitted for publication, IEEE Trans. SMC Part B. January 2014.

• W. A. Albukhanajer, Y. Jin, and J. A. Briffa, “Neural Network Ensembles for Image

Identification Using Pareto-optimal Features,” in 2014 IEEE Congress on Evolutionary Computation (CEC), Beijing, China, 06-11, July 2014.

• W. A. Albukhanajer, J. A. Briffa, and Y. Jin, “Classifier Ensembles Using Pareto Optimal

Image Features ” To be submitted, IEEE Trans. Image Proc. June 2014.

Wissam A. Albukhanajer

w.albukhanajer@surrey.ac.uk

Nature Inspired Computing and Engineering

T: +44(0)1483 68 6059 F: +44(0)1483 68 6051

Thank you

Q&A

Department of Computing Faculty of Engineering & Physical Sciences University of Surrey Guildford, UK. GU2 7XH

http://www.surrey.ac.uk/computing/ 7th July 2014