GAdaBoost: Accelerating Adaboost Feature Selection with Genetic

Algorithms

Mai F. Tolba and Mohamed Moustafa

Computer Science and Engineering Department, The American University in Cairo, Road 90, New Cairo, Cairo, Egypt

Keywords:

Object Detection, Genetic Algorithms, Haar Features, Adaboost, Face Detection.

Abstract:

Boosted cascade of simple features, by Viola and Jones, is one of the most famous object detection frame-

works. However, it suffers from a lengthy training process. This is due to the vast features space and the

exhaustive search nature of Adaboost. In this paper we propose GAdaboost: a Genetic Algorithm to acceler-

ate the training procedure through natural feature selection. Speciﬁcally, we propose to limit Adaboost search

within a subset of the huge feature space, while evolving this subset following a Genetic Algorithm. Experi-

ments demonstrate that our proposed GAdaboost is up to 3.7 times faster than Adaboost. We also demonstrate

that the price of this speedup is a mere decrease (3%, 4%) in detection accuracy when tested on FDDB bench-

mark face detection set, and Caltech Web Faces respectively.

1 INTRODUCTION

Machine learning and training, require large feature

sets, which can be time consuming to explore. With

the advancement of this ﬁeld the need for algorithms

to decrease the training time arose. Genetic Algo-

rithms (GA) has proven its strength in solving prob-

lems like the aforementioned one, especially those

concerned with exploring large search spaces, and

providing acceptable results in a signiﬁcantly less

amount of time than the brute force manner. Many re-

search has explored the use of GA in time consuming

tasks like feature selection, which aims to choose a

representative small subset of features from the whole

set of features (Xue et al., 2015).

Object detection lies in the set of machine learn-

ing techniques that require a huge search space for

training, thus their training is time consuming. Ob-

ject detection is concerned with detecting whether an

object is present in a given image, and where it lies

in this image. It has many applications including, but

not limited to, face detectors in all modern state of the

art cameras, automotive safety, video indexing, image

classiﬁcation, surveillance, and content-based image

retrieval (Lillywhite et al., 2013).

Much research has been put into this area, due to

its complex nature as detection is hard to achieve in

different light conditions, occlusion and the angle in

which the object appears in the image (Lienhart and

Maydt, 2002; Lillywhite et al., 2013; Viola and Jones,

2001). Researchers has been trying to implement ef-

ﬁcient high speed detectors that work in real time

and has a high percentage of accuracy. Though the

Viola-Jones detector has reached an impressive de-

tection speed it still consumes a lot of time in train-

ing. Viola-Jones uses Adaboost, a type of boosting

algorithms, to select and combine weak classiﬁers to

form a strong one. Adaboost is simple and adaptive

(Dezhen and Kai, 2008), yet it operates in a brute

force manner, passing by all the set of features mul-

tiple times. This can be very time consuming, as the

search space consists of a set of more than 162,000

features for a 24X24 image.

The main contribution of this paper is increasing

the speed of training of the Viola-Jones face detector

by implementing a hybrid approach that combines the

use of boosting and GAs. This will ﬁnd the best fea-

tures efﬁciently instead of going through all of them

in a brute force manner, which will decrease the train-

ing time. The Paper is organized as follows: section 2

discusses some of the related previous work. Section

3 explains the proposed method. Section 4 provides

some experiments and their results. Section 5 con-

cludes the paper and discuss future enhancements.

2 RELATED WORK

Genetic Algorithms (GA) are optimizing procedures

that are devised from the biological mechanism of

156

Tolba, M. and Moustafa, M.

GAdaBoost: Accelerating Adaboost Feature Selection with Genetic Algorithms.

DOI: 10.5220/0006041101560163

In Proceedings of the 8th International Joint Conference on Computational Intelligence (IJCCI 2016) - Volume 1: ECTA, pages 156-163

ISBN: 978-989-758-201-1

reproduction, and evolutionary science (Sun et al.,

2004). GA continues to prove itself successful in

many ﬁelds including object detection. There are

other optimization methods that serve well; however,

from some experiments, GA has proven to perform

better in solving problems. This might be due to the

advantages of GA and that they are probabilistic and

not deterministic, and have the ability to be better at

avoiding to be stuck at a local maxima and are par-

allelizable. Ferri et al. (1994), compared GA against

sequential search and their results clearly show that

GA performs better. Their work highlighted the point

of strength of GA which is the ability to perform the

search in a near optimal region due to the inherited

randomizations used in the search. Tabassum and

Mathew (2014), said that It was proved that genetic

algorithms are the most powerful unbiased optimiza-

tion techniques for sampling a large solution space.

After implementing the Knapsack problem and im-

age optimization, they concluded in their paper that

GA are the best application to solve various common

problems and that they are suitable for solving high

complexity problems like the combinatorial optimiza-

tions. Sun et al. (2004) provided another proof to the

strength of GA, when they used it to select the best

eigenvectors. In their work GA was used to solve the

problem of selecting the best feature set. They com-

pared their results with other techniques and proved

to provide better accuracy with less number of fea-

tures. Lillywhite et al. (2013) used Genetic algo-

rithms in constructing features which was used by

Adaboost to train a classiﬁer. They tested their ap-

proach against previously published papers and used

the same dataset for comparison. Their technique

proved to be signiﬁcantly more accurate than most of

the previous work they compared against. Some re-

searchers used GA in feature selection. Feature selec-

tion methods can be divided into 3 main categories:

wrappers, ﬁlters and embedded methods. Filters are

a form of preselecting each feature on its own with-

out considering the previous predictor. Wrappers are

methods to score the predictive power of a subset of

features by using a machine learning technique as a

black box, while the embedded are methods that inte-

grate classiﬁcation and learning into a single process

(Chouaib et al., 2008; Xue et al., 2015) . In Their

work, Chouaib et al. (2008) aimed to ﬁnd the set of

the most representative features using GAs, in order

to decrease the detection time. Their results showed

that for the majority of descriptors their feature set

was signiﬁcantly reduced up to 75% of the original

set in two class problems. Dezhen and Kai (2008)

provided a post optimization technique to avoid the

redundancy of classiﬁers. By doing so, they managed

to increase the speed of classiﬁcation by 110% due to

reducing the number of features to 55% of the original

set. Since this is a post optimization process, it can be

considered as an added part to the training process,

which will be an overhead on the training time. Xue

et al. (2015) provided a survey on the use of evolu-

tionary computing in feature selection. In their work

they surveyed more than 40 papers which use GA in

feature selection.

Object detection is a main area of research in com-

puter vision. It falls under the type of problems that

suffer from a time consuming training process, due

to the huge search space involved. Viola and Jones

(2001) devised a new face detector using Haar fea-

tures, since features provide a set of comprehensive

information that can be learned by machine learning

algorithms. They also reduce the in-out class vari-

ability compared to that of the raw pixels (Lillywhite

et al., 2013; Viola and Jones, 2001). Haar features

are mainly rectangles divided into black and white re-

gions and the value of this feature is calculated by

subtracting the sum of pixels in the white region from

the sum of those in the black region (Viola and Jones,

2001). For each image, variations of each of the four

Haar feature types are computed in all possible sizes

and all possible locations, which provides a huge set

of features.

The authors chose Adaboost as a method to obtain

their strong classiﬁer. Adaboost was proposed by Fre-

und and Schapire (1995), it has the power to search

through the features and select those of good perfor-

mance then combine them to create a strong classiﬁer.

The general idea of the algorithm works as follows:

For a number of iterations T:

• Pass through the set of all possible features and

calculate the error of each one on the given im-

ages.

• Choose the best feature (the one with the lowest

error) as the ﬁrst weak classiﬁer.

• Update the sample images corresponding weights,

by putting more weights on the wrongly classiﬁed

images.

• It then goes through the next iteration, until it

ﬁnds the set of best features, to be used in clas-

siﬁcation.

One of the important contributions of Viola and

Jones (2001) work is the cascade classiﬁer which in-

creased the accuracy while radically reducing the time

consumed in detection. The cascade classiﬁer is a

stage classiﬁer where the thresholds vary. The ﬁrst

stages has a low threshold, thus detecting all the true

positives while eliminating the strong negatives, be-

fore more complex classiﬁers are used to achieve less

GAdaBoost: Accelerating Adaboost Feature Selection with Genetic Algorithms

157

false positives. Although their ﬁnal detector has per-

formed well in terms of accuracy, the training process

is time consuming as Adaboost passes by the set of all

possible features multiple times in training each stage

of the classiﬁer. In addition to that, training the stage

classiﬁer becomes slower in the last stages since the

images become harder to classify. This is due to the

use of images that were classiﬁed as false positives

by the previous stages, therefore the last stages of the

classiﬁer typically need more features to satisfy the

false alarm and the true positive rates desired. This

means that the ﬁnal stages has to go through the set of

all possible features more times than in the ﬁrst stages

which puts an overhead on the training time.

Some of the researchers used the Viola-Jones al-

gorithm as a base for their research to provide a more

powerful detector. Lienhart and Maydt (2002) pro-

posed the increase of the Haar features used. Instead

of using only the four basic types of Haar features,

they increased them to include the 45 degree rotation

of the original ones, consequently, the total number

of feature types amounted to 14. The use of more fea-

tures resulted in better accuracy, yet it substantially

increased the number of generated features per im-

age. A larger feature set means that more time will

be taken by the training as the Adaboost operates in a

brute force manner.

The previous work was concerned with enhanc-

ing the accuracy or speed of detection regardless of

the overhead posed on the training time. This work

aims to examine the effects on increasing the speed

of training using GA and how this might affect the

accuracy.

3 PROPOSED METHOD

The proposed method (Named: GAdaBoost) applies

GA to select a set of features, to have Adaboost

choose from, instead of going through the set of all

possible features. The original Adaboost algorithm

was proposed by Freund and Schapire (1995) the gen-

eralized version works as follows: for the training of

each stage in the stage classiﬁer, the algorithm passes

through the set of all possible features and calculates

the error of each feature on each given image. Af-

ter that, it chooses the best feature (the one with the

lowest error, i.e best classiﬁes the image correctly) as

the ﬁrst weak classiﬁer. It then updates the sample

images and their corresponding weights, by putting

more weights on the wrongly classiﬁed images. The

procedure is repeated until the set of chosen features

reaches a preset false alarm, and hit rate set for clas-

siﬁcation.

In order to integrate the use of GA, OpenCV

s (It-

seez, 2015) implementation on the Viola Jones algo-

rithm has been modiﬁed. Incorporating the use of GA

will increase the training speed by avoiding the error

calculation of the set of all possible features, and only

providing the Adaboost algorithm with a representa-

tive set of features, that have been chosen based on

their classiﬁcation power. This set of representative

candidate features is to be prepared by the GA before

the training of each stage in the ﬁnal classiﬁer. For

example if the ﬁnal classiﬁer is to have 10 stages the

added GA technique is to be repeated 10 times. The

stage training utilizes Adaboost technique to choose

multiple weak classiﬁers from the mentioned repre-

sentative set, in order to reach the desired false alarm

and hit rate preset for the stage. Figure 1 shows a

block diagram that explains the proposed GAdaBoost

technique.

Figure 1: GAdaBoost diagram.

On the ﬁrst iteration the GAdaboost chooses a pre-

set number of features randomly to create the ﬁrst

generation of the given population size. Those ran-

domly chosen features are marked so that they are not

to be used again when more random features are to

be generated. This is done to explore more of the

set of all possible features. In order to assess the

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

158

predictive power of these features, they are passed

to a learning algorithm. The way this has been im-

plemented is by creating a temporary (dummy) stage

where the features are trained in the same way the

original stage training works, i.e the dummy stage is

an Adaboost training algorithm. The number of weak

classiﬁers chosen by the Adaboost algorithm in the

dummy stage is a variable that is preset. The Ad-

aboost algorithm associates the features with scores

that are a representation of their predictive power. Af-

ter that the best features are then selected and have

mutation and crossover processes preformed on them

to get the next generation of an even better preforming

set of features. The new generation is then passed by

a dummy stage for scoring. The process is repeated

until the average ﬁtness of the population saturates

or a predeﬁned number of iterations is reached. As

a form of exploring more of the set of all possible

features, for each iteration with an even number (2nd,

4th, etc. generations) that is greater than zero, the best

set of parents and their children produced are chosen.

Then a spatial comparison is formed to remove the

redundant features, and random features are inserted

instead to complete the population size. The spatial

comparison is done using the pasacal criterion where

two features are considered of spatial similarity if the

ratio of the intersection of the two features over the

union of the two is greater than 0.4. This method is

described in more details in section 4.3. The use of

only even iterations entails that the spacial compari-

son is done on half the number of iterations (eg for 50

iterations, the spatial comparison is done 25 times).

The ﬁnal set of features obtained by the GA is passed

through a real stage where the weak classiﬁers se-

lected by this stage are to be used in the resultant ﬁnal

classiﬁer. The afore mentioned technique ensures that

the Adaboost algorithm will only evaluate the popu-

lation size chosen instead of going through the whole

set of features when selecting the weak classiﬁers of

the resultant ﬁnal stage classiﬁer.

As for the Genetic Algorithm parameters, each

chromosome represents one Haar feature. The val-

ues of the chromosome are x, y, x1, y1, type, where x,

y are the upper left coordinates of the feature rectan-

gle, and x1, y1 are the lower right corner. The type

is a value from 0-4 where each number represents

one of the Haar Feature types used for upright frontal

faces detection. 0,1,2,3,4 represent the Haar types of

haar x2, haar y2, haar x3, haar y3, and haar x2y2 re-

spectively. Figure 2 explains the mapping of a feature

of type haar x2 to a chromosome in a given image.

As shown the chromosome carries decoded informa-

tion about the type of the feature and its orientation

in a given image, the way the chromosome is repre-

Figure 2: Chromosome to feature mapping.

sented facilitates the mutation and the crossover pro-

cesses which provides new features. The ﬁtness func-

tion is a measure of how well this features splits be-

tween the negative and the positive images, or in other

words predictive power of this feature in classifying

the images correctly. The OpenCV implementation

uses decision stumps as weak classiﬁers, these deci-

sion stumps are Classiﬁcation and Regression trees

(CART). In CART the regression trees best split qual-

ity is calculated by the minimization of Equation 1.

∑

i=1

(T R

− PR

)

(1)

Where TR is the ground truth of the image, PR

is the predicted responses by the decision stump and

n is the number of sample images. Yet, for simplicity

the OpenCV traincascade developer mentioned that in

implementation the minimization criteria is reduced

to equivalent simpler maximization ones (Dimashova,

2012). The ﬁtness of the feature used is the split qual-

ity measure provided by OpenCV

s CVDTree class.

Thus, in the implementation the best feature is the

one the largest quality. The candidates to be used for

crossover and mutation are selected using a Roulette

Wheel selection method, so that those features with

higher ﬁtness are more likely to be selected and have

the mutation and the crossover applied on them. A

simple one-point crossover is utilized at the lower

right corner of the two candidate features. In order

to reduce the time taken by validation of the correct-

ness of the feature, the mutation is designed to assign

the type to the feature according to how suitable this

type is, given the coordinates of this feature.

GAdaBoost: Accelerating Adaboost Feature Selection with Genetic Algorithms

159

4 EXPERIMENTS AND RESULTS

In this section we discuss our experimental work, its

setup, and the acquired results. It is divided into 3

subsections, each one describes an experiment setup

and discusses the results acquired. The ﬁrst section

shows the ﬁtness of the individual and the average

ﬁtness of the population. The second one shows the

effects of varying the population size on the training

time. The ﬁnal one shows the effect of varying the

number of iterations of GAdaBoost and comparing it

against the original brute force algorithm with respect

to the training time and accuracy. The testing was

performed on both the FDDB dataset, and Caltech

10,000 Web Faces dataset. All the training occurred

on the same computer with an Intel Core i7-4510U @

2.00GHz processor and 8 GB RAM.

The positive images used for training are the im-

ages provided by OpenCV for the upright frontal

faces training. The negative images were picked ran-

domly from the dataset of 101 objects developed at

Caltech (Fei-Fei et al., 2004). All the experiments

use the same settings of a 17 stage cascade classiﬁer

with 500 positive images and 500 negative images per

stage, hit rate of 0.9 and a false alarm rate of 0.5 per

stage. Each dummy stage is trained for 3 weak clas-

siﬁers and the sample image weights are carried on

between the dummy stages, with the exception of ex-

periments in section 4.1 where each dummy stage has

been trained for only one weak classiﬁer, with no car-

rying on of the image weights between dummy stages,

and the check on the special proximity and its removal

wasn

t utilized.

4.1 Individual and Population Fitness

In this Experiment the GAdaBoost discussed in the

proposed method (section 3) has been used to train a

cascade classiﬁer with the settings mentioned in the

beginning of this section. A population size of a 1000

and 50 iterations were the parameters set for the GAd-

aBoost. Figure 3 shows the progress of the best indi-

vidual, and the average ﬁtness of the population. They

are shown over the course of the 50 iterations of the

GA preformed before the 17th stage.

As shown from Figure 3, the average population

ﬁtness increases fast in the ﬁrst iterations then it starts

to saturate, while the ﬁtness trend of best individual is

to increase then saturate and so on.

4.2 Population Size vs Training Time

In this Experiment 20 iterations is set for GAdaboost.

Each classiﬁer has been trained multiple times and

Figure 3: Best individual ﬁtness and average population ﬁt-

ness over 50 iterations.

the average time taken has been calculated. Figure

4 shows the effect of varying the population size on

the training time.

From Figure 4 it is clear that the training time

increases as the population size increases. This

can be attributed to the fact that less mutations and

crossovers are done with a smaller population size,

and that the GA provides the Adaboost with a smaller

feature set to go through in a brute force manner.

Figure 4: Population size vs training time.

4.3 Training Speed Up vs. Accuracy

As a baseline a cascade classiﬁer has been trained

using OpenCVs traincascade method using the same

settings. This classiﬁer has been used as a base-

line to compare our algorithm with. We performed

our experiment on both the Face Detection Dataset

(FDDB)(Jain and Learned-Miller, 2010), and Caltech

10,000 Web Faces dataset (Angelova et al., 2005).

Face detection datasets were chosen since Haar fea-

tures were used originally to detect upright frontal

faces. The FDDB dataset is a benchmark dataset de-

signed for studying the unconstrained face detection

problem. This dataset has been used in many studies

and is considered one of the difﬁcult datasets, due to

occlusions, out of focus faces and difﬁcult poses (Jain

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

160

and Learned-Miller, 2010). This dataset contains an-

notations for 5171 faces in a set of 2845 images. In

this paper 290 images from this dataset has been used

for experimentation. For the evaluation of our detec-

tor on this dataset, we used the evaluation code pro-

vided by the authors of this dataset.

Caltech Web Faces is a dataset of human faces col-

lected from the web (Angelova et al., 2005). It is a

challenging dataset since it contains difﬁcult exam-

ples such as extreme face orientations, occlusion like

hats and glasses and variable light conditions (An-

gelova et al., 2005). The Caltech Web Faces data

set consists of 10,524 annotated faces where the co-

ordinates of the mouth, eyes and nose are given. For

our experiments we randomly extracted 500 images

to test on. We used the given eye coordinates to set

a bounding square by assuming that the eye distance

is half the face width. For evaluation, we use the pas-

cal criteria shown in Equation 2 (Everingham et al.,

2010).

area(B

∩ B

det

)

area(B

∪ B

det

)

> 0.4 (2)

Where B

is the ground truth bounding box and

det

is the detected bounding box. Thus the ratio of

the area of intersection between the two boxes to the

area of their union has to exceed 0.4 in order for the

detected box to be counted as a face.

To test our algorithm we examined two variations

of the GAdaBoost cascade classiﬁer where we vary

the number of iterations. GAdaboost has been trained

using 20 and 50 iterations. Every experiment has been

run multiple times, and an average of the training time

and the performance of all the runs has been calcu-

lated. The speed of training vs accuracy of the two

GAdaboost variations have been compared against

the baseline.

The obtained results show that training GAd-

aBoost in the two experiments took signiﬁcantly less

time than training the baseline. Training our approach

with 20 and 50 iterations took about 26.8% and 44.7

% percent of the time taken to train the baseline re-

spectively. Figure 5 visually emphasizes the differ-

ence in training time taken by the baseline and both

algorithms.

Figures 6 and 7 provide the Y error bar graphs,

showing the maximum, minimum and average results,

for all the runs of both the 20 and the 50 iterations

GAdaBoost on FDDB dataset. Its clear from the ﬁg-

ures that the 50 iterations GAdaboost preforms better.

It even has best case scenarios where the accuracy was

almost the same as the baseline, for certain numbers

of false positives.

From both Figures 6 and 7, by examining the av-

Figure 5: Training time in minutes of each of the experi-

ments.

Figure 6: Y error bars for all the runs of the 20 iterations

GAdaBoost on FDDB.

Figure 7: Y error bars for all the runs of the 50 iterations

GAdaBoost on FDDB.

erage point on the Y error bars it can be observed

that at 500 false positives the baseline true positive

rate is 64% and the GAdaBoost 20 and 50 iterations

achieved 58% 59% true positive rate respectively.

While at 1000 false positives the baseline achieved

GAdaBoost: Accelerating Adaboost Feature Selection with Genetic Algorithms

161

67% true positive rate versus about 64% and 65% for

the GA 20 and 50 iterations respectively. Collectively

from the provided ﬁgures, it can be noted that GAd-

aBoost with 50 iterations has performed slightly bet-

ter than the GAdaBoost with 20 iterations. It can also

be observed that at lower thresholds the GA provides

closer true positive rates compared with the baseline,

than it does at higher thresholds.

Figures 8 and 9 show the Y error bar graphs,

showing the maximum, minimum and average results,

for all the runs of both the 20 and the 50 iterations

GAdaBoost on Caltech Web Faces dataset.

Figure 8: Y error bars for all the runs of the 20 iterations

GAdaBoost on Caltech Web Faces.

Figure 9: Y error bars for all the runs of the 50 iterations

GAdaBoost on Caltech Web Faces.

From both Figures 8 and 9, by examining the av-

erage point on the Yerror bars we ﬁnd that at 500 false

positives the baseline true positive rate is 46 % and the

GAdaboost 20 and 50 iterations achieved 41% , 43%

true positive rate respectively. While at 1000 false

positives the baseline achieved 51% true positive rate

versus about 47% and 48% for the GAdaBoost 20 and

50 iterations respectively. Collectively from the pro-

vided ﬁgures, it can be noted that GAdaboost with

50 iterations has performed slightly better than the

GAdaboost with 20 iterations. The decrease in per-

formance of both the baseline and GAdaboost can be

attributed to the fact that Caltech Web Faces dataset

includes occlusions and light variations as was men-

tioned at the beginning of this section.

5 CONCLUSIONS

We showed the effect of incorporating Genetic Algo-

rithms with the Viola-Jones Rapid object detector on

enhancing the training speed. Experiments to show

the progression of the best individual and the average

population ﬁtness were provided. Other experiments

showed the speed up that can be gained by the re-

duction of the population size. Also, two variations

of the GAdaboost were examined, one with 20 iter-

ations and the other with 50 iterations. Both exper-

iments were run multiple times to observe the effect

of the number of iterations on the performance using

the FDDB and Caltech Web Faces dataset. We ex-

perienced that the training process became up to 3.7

times faster than the original algorithm with a mere

decrease of 3% to 4% in accuracy. We noted that the

50 iterations performed better than the 20 iterations,

and both had best case scenarios of almost reaching

the baseline accuracy at some thresholds.

The future extension of this contribution can be

done by experimenting with more GAdaboost param-

eters by varying the iteration numbers, or ﬁnding bet-

ter stopping criteria for the GA. The parallelizable na-

ture of the GA can be utilized to gain an even faster

training process.

REFERENCES

Angelova, A., Abu-Mostafa, Y., and Perona, P. (2005).

Pruning training sets for learning of object categories.

Proceedings of the IEEE Computer Society Confer-

ence on Computer Vision and Pattern Recognition,

1:494–501.

Chouaib, H., Terrades, O. R., Tabbone, S., Cloppet, F.,

and Vincent, N. (2008). Feature selection combin-

ing genetic algorithm and adaboost classiﬁers. In

Pattern Recognition, 2008. ICPR 2008. 19th Interna-

tional Conference on, pages 1–4.

Dezhen, Z. and Kai, Y. (2008). Genetic algorithm based

optimization for adaboost. In Computer Science and

Software Engineering, 2008 International Conference

on, volume 1, pages 1044–1047.

Dimashova, M. (2012). How is decision tree split quality

computed. http://answers.opencv.org/question/566/

how-is-decision-tree-split-quality-computed/.

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

162

Everingham, M., Gool, L. V., Williams, C. K. I., and Winn,

J. (2010). The P ASCAL Visual Object Classes ( VOC

) Challenge. International Journal, pages 303–338.

Fei-Fei, L., Fergus, R., and Perona, P. (2004). Learning

Generative Visual Models From Few Training Exam-

ples: An Incremental Bayesian Approach Tested on

101 Object Categories. In IEEE CVPR Workshop of

Generative Model Based Vision (WGMBV).

Ferri, F., Pudil, P., Hatef, M., and Kittler, J. (1994). Com-

parative study of techniques for large-scale feature se-

lection*. In GELSEMA, E. S. and KANAL, L. S.,

editors, Pattern Recognition in Practice IVMultiple

Paradigms, Comparative Studies and Hybrid Sys-

tems, volume 16 of Machine Intelligence and Pattern

Recognition, pages 403 – 413. North-Holland.

Freund, Y. and Schapire, R. (1995). A desicion-theoretic

generalization of on-line learning and an application

to boosting. Computational learning theory, 55:119–

139.

Itseez (2015). Open Source Computer Vision Library.

https://github.com/itseez/opencv.

Jain, V. and Learned-Miller, E. (2010). FDDB: A Bench-

mark for Face Detection in Unconstrained Settings.

Technical Report UM-CS-2010-009, University of

Massachusetts, Amherst.

Lienhart, R. and Maydt, J. (2002). An extended set of

Haar-like features for rapid object detection. Proceed-

ings. International Conference on Image Processing,

1:900–903.

Lillywhite, K., Lee, D. J., Tippetts, B., and Archibald, J.

(2013). A feature construction method for general ob-

ject recognition. Pattern Recognition, 46(12):3300–

3314.

Sun, Z., Bebis, G., and Miller, R. (2004). Object detection

using feature subset selection. Pattern Recognition,

37(11):2165–2176.

Tabassum, M. and Mathew, K. (2014). A Genetic Algo-

rithm Analysis towards Optimization solutions. Inter-

national Journal of Digital Information and Wireless

Communications (IJDIWC), 4(1):124–142.

Viola, P. and Jones, M. (2001). Rapid object detection using

a boosted cascade of simple features. Proceedings of

the 2001 IEEE Computer Society Conference on Com-

puter Vision and Pattern Recognition. CVPR 2001, 1.

Xue, B., Zhang, M., Member, S., and Browne, W. N.

(2015). A Survey on Evolutionary Computation Ap-

proaches to Feature Selection. 2007(September):1–

20.

GAdaBoost: Accelerating Adaboost Feature Selection with Genetic Algorithms

163