연령 추정 알고리즘은 계층적 접근 방식을 실현합니다(그림 10). 첫째, 입력 조각은 18세 미만, 18-45세, 45세 이상의 세 연령 그룹으로 나누어 집니다. 둘째, 이 단계의 결과는 각각 10년 단위로 제한되는 7개의 작은 그룹으로 세분화됩니다. 따라서, 다중 클래스 분류 문제는 "일대다" 이진 분류자 집합으로 감소합니다. 따라서 이 분류자는 관련 클래스를 기반으로 이미지의 순위를 매기고, 이러한 순위 히스토그램을 분석하여 최종 결정을 내립니다.
이 BC들은 2단계 접근방식을 사용하여 구성됩니다. 앞에서 설명한 대로 이미지는 적응형 특징 공간으로 처음 전환한 후 RBF 커널이 있는 지원 벡터 머신을 통해 분류됩니다.
입력 조각은 밝기 특성이 균일한 척도로 정렬되고 변환되도록 사전 처리됩니다. 이 사전 처리 단계에는 색상 공간 변환 및 스케일링이 포함되며, 두 작업 모두 성별 인식 알고리즘에 사용된 과정과 유사합니다. 특징은 각 색상 구성 요소에 대해 계산되고 결합되어 균일한 특징 벡터를 형성합니다.
교육 및 테스트에는 충분히 큰 색상 이미지 데이터베이스가 필요합니다. 우리는 서로 다른 출처에서 얻은 10,500개의 얼굴 이미지를 구성하는 자체 이미지 데이터베이스와 최첨단 MORF 및 FG-NET 이미지 데이터베이스를 결합했습니다. AdaBoost 얼굴 감지 알고리즘이 이미지의 얼굴을 자동으로 감지합니다.
연령 분류 알고리즘의 첫 번째 단계를 훈련하고 테스트하는 데 총 7000개의 이미지가 사용되었습니다. 144개의 적응형 특징을 사용하여 3개의 BC가 생성되었습니다.
1단계 분류 결과는 젊은 얼굴의 경우 82%의 정확도를, 중년의 얼굴은 58%, 노년의 얼굴은 92%의 정확도를 보였습니다. 세 가지 연령 범주에 대한 전반적인 연령 분류 정확도는 77.3%였습니다.
두 번째 단계의 BC는 첫 번째 단계(위에서 설명)와 동일한 방식으로 구성되었습니다. 그림 11은 제안된 알고리즘의 첫 번째 단계에 의한 연령 추정의 예를 시각적으로 보여줍니다.
The age estimation algorithm realises hierarchical approach (fig. 10). First of all input fragments are divided for three age groups: smaller than 18 years old, from 18 - 45 years old and bigger than 45 years old. Afterwards the results of this in the first stage are more subdivided to seven newer groups with each limiting to one single decade. Thus the problem of multiclass classification is therefore reduced to a set of binary ‘one-against-all’ classifiers (BC).These classifiers calculate: ranks of each of the analyzed class. The total decision is obtained then by the analysis the previously received histogram of ranks.
These BCs are construction is applied with the transitioning to adaptive feature space, equal to this described earlier, and support vector machines classification of images with RBF kernel.
Input fragments were preprocessed for their luminance characteristics to align and to transform them to uniformal scale. Preprocessing includes color-space transformation and scaling, both similar to that of gender recognition algorithm. Features, calculated for each colour component, are combined to form a uniform featured vector.
Training require a huge enough coloring image database: We used state-of-the-art image databases MORPH and FG-NET with our own image database, gathered from some sources, which comprised of 10500 faces. Faces on the images were detected automatically by AdaBoost face detection algorithms.
A total number of seven thousand images were used for age classification algorithm training and testing on the first stage. 3 binary classifiers were made utilizing 144 adaptive features each of.
Classification results on the first stage are: 82 % accuracy for young age, 58 % accuracy for middle age and 92 % accuracy for senior age. Age classification rate in a three age division problem – 77.3 %.
Binary classifiers of the second stage were constructed equal to the first stage described above. A visual example of age estimation by the proposed algorithm on its first stage is presented in figs. 11.
The age
estimation algorithm realises hierarchical approach (fig. 10). First of all input
fragments are divided into1for three age groups: smaller than 18 years old,
from 18 - 45 years old and bigger than 45 years old. Afterwards the results of this
in the first stepage are more subdivided to seven smallernewer
groups, with
each limiteding
to one single decade. Thus the problem of multiclass classification is therefore
reduced to a set of binary ‘one-against-all’ classifiers (BC).These classifiers
calculate: ranks of each of the analyzed class. The total decision is obtained then
by the analysis the previously received rank histograms of ranks.
These BCs are
constructioned
using a two-level approach. After is appliedfirst with the 2transitioning to adaptive feature space,
asequal to this3
described earlier, and support vector machines classification of images with
RBF kernel.
Input fragments were
preprocessed for their luminance characteristics to align and to transform them
to uniformal scale. This pPreprocessing step includes
color-space transformation and scaling, both operations
similar to those used inthat of a gender
recognition algorithm. Features, calculated for each colour
component, are combined to form a uniform featured4
vector.
Training and testing5 require a sufficientlyhuge largeenough coloring
image database: We used state-of-the-art image databases MORPH and FG-NET with our
own image database, gathered from some6many sources, which comprised of 10,500 face images.
Faces on the images were detected automatically by AdaBoost face detection
algorithms.
A total number
of seven thousand images were used for age classification algorithm training
and testing on the first stage. 3 binary classifiers were made utilizing 144
adaptive features each of.
Classification
results on the first stage are: 82 % accuracy for young age, 58 % accuracy for
middle age and 92 % accuracy for senior age. Age classification accuracyrate
in a three age division problem – 77.3 %.
Binary classifiers
of the second stage were constructed equalsame to the first stage described above. A visual
example of age estimation by the proposed algorithm on its first stage is
presented in figs. 11.
The proposed age estimation algorithm realisesrealizes
hierarchical approach (figFig. 10). First of
all, the input fragments are divided into1for three
age groups: smallerless than 18 years old, from 18 - 45 years old and biggermore than 45 years old. Afterwards2Second,
the results of this in the firstfirst stepage are more further subdivided
tointo
seven smallernewer
groups, with
each limiteding
to onea
single decade. ThusThis reduces the problem
oforiginal multiclass
classification is therefore reduced problem
to a set of binary ‘”one-against-all’”3 classifiers (BC). These
classifiers calculate:Each classifier
then ranks of each of the analyzedthe images based on the associated class. The total decision is, and the final decisions are obtained then by the analysis
the previously received analyzing these rank histograms of ranks.
These BCs are
constructioned4
using a two-level approach. After is appliedfirst with the transitioning to an adaptive feature space, asequal to this5
described earlier, and the images are classified using support vector
machines classification of images with radial basis function (RBF) kernels.6
The Iinput fragments wereare preprocessed
forto align
and transform their luminance characteristics to align and to transform them to uniformala uniform scale. This
pPreprocessing step includes color-space transformation and
scaling, both operations similar to those used inthat of
athe 7gender
recognition algorithm. Features, are calculated for
each colourcolor
component, are and combined to form a uniform featured8 vector.
Training and testing9 require a sufficientlyhuge largeenough coloring
image database: We used. Here, we combined the state-of-the-art image databases MORPH and FG-NET image databases with our own image database,
gathered from some10many sources, which comprised of many sources and comprising 10,500 face images. Faces on. The faces
in the images were detected automatically by the AdaBoost face detection algorithms.
A total number of seven thousand7000 images were used for
to train and test the first stage of the
age classification algorithm training and testing
on the first stage. 3 binary classifiers.
Three BCs were made utilizingcreated, each with11 144 adaptive features each of.
Classification
results on the The first-stage are:classification results showed 82 % accuracy for
young agefaces,
58 % accuracy for middle age-aged faces, and 92 % accuracy for seniorelderly faces. The overall
age. Age classification accuracyrate
in afor the
three age division problem –categories was 77.3 %.12
Binary
classifiers of tThe
second-stage BCs
were constructed in the equalsame tomanner as the
first stage (described above. A). Fig. 11 shows a
visual example of age estimation by the first stage
of the proposed algorithm on its first
stage is presented in figs. 11..
The proposed age estimation algorithm realizes hierarchical approach (Fig. 10). First, the input fragments are divided into three age groups: less than 18 years old, 18–45 years old, and more than 45 years old. Second, the results of this first step are further subdivided into seven smaller groups, each limited to a single decade. This reduces the original multiclass classification problem to a set of binary “one-against-all” classifiers (BC). Each classifier then ranks the images based on the associated class, and the final decisions are obtained by analyzing these rank histograms.
These BCs are constructed using a two-level approach. After first transitioning to an adaptive feature space, as described earlier, the images are classified using support vector machines with radial basis function(RBF) kernels.
The input fragments are preprocessed to align and transform their luminance characteristics to a uniform scale. This preprocessing step includes color-space transformation and scaling, both operations similar to those used in the gender recognition algorithm. Features are calculated for each color component and combined to form a uniform feature vector.
Training and testing require a sufficiently large color image database. Here, we combined the state-of-the-art MORPH and FG-NET image databases with our own image database, gathered from many sources and comprising 10,500 face images. The faces in the images were detected automatically by the AdaBoost face detection algorithms.
A total of 7000 images were used to train and test the first stage of the age classification algorithm. Three BCs were created, each with 144 adaptive features.
The first-stage classification results showed 82% accuracy for young faces, 58% accuracy for middle-aged faces, and 92% accuracy for elderly faces. The overall age classification accuracy for the three age categories was 77.3%.
The second-stage BCs were constructed in the same manner as the first stage (described above). Fig. 11 shows a visual example of age estimation by the first stage of the proposed algorithm.
