| Summary: | Age estimation through face image has a lot of potential computer applications, such as
social media communication and access control. However, it is a challenging problem
for the existing methods in computer systems to effectively estimate humanfacial age.
There are many methods for age estimation, but they are significantly lacking in their
performances, especially when compared to other related task performances in face
recognition. Not only the reflection of face structure and face texture is incomplete, the
complexity of feature has not been properly addressed. For this reason, facial features
need to be highly compressed to avoid a high degree of complexity while
accommodating larger users and increasing data relation. The objectives of this study
are to develop a facial feature extra tion model using machine learning based
approaches for extracting features from a human face and to map the features to an
output label that estimates the predicted age (in years) of an individual. An integrated
facial feature extraction model was de eloped which combines the power of both
models, namely Active Appearance Mod 1 (AAM) and Active Shape Model (ASM), in
order to obtain an innovative result on the problem of feature detection. A two-step
app oach for combining AAM and ASM was suggested. In the first step, the ASM was
utilized to locat the outer shap Ian marks of the face, and in the second step, AAM
was utilized to locate the inner shape landmarks of the face. The experiments were
tested on the publicly accessible MORPH and LFBW databases. The dataset consists of
490 training images and 210 test images per database. The developed algorithm was
evaluated by measuring the performance of the facial feature extraction with respect to
feature accuracy and error rates. A few prominent machine learning algorithms were
then explored for age prediction. A hierarchical approach was developed to train the
machine learning algorithms for facial features with age labels. Secondly, an
expectation framework was developed to jointly address the issues of categorizing the
test features into age labels based on the trained data. The performance was evaluated
using Mean Absolute Error (MAE), Cumulative Score (CS) and processing time. The
experimental results indicated that the proposed model is effective in extracting facial
features and estimating age. For LFPW and MORPH databases, the error rates were
2.2628 and 2.7174 respectively. The best MAE value obtained from the proposed model
is 2.94, achieved in Canonical Correlation Analysis (CCA). The best CS value obtained
from the proposed model is 88.9%, achieved in CCA. The best processing time obtained
from the proposed model is 0.036142, achieved in Support Vector Machine (SVM).
These results are more promising when compared to several states of the art techniques
like ASM and AAM, in which the error rates were 2.6607 and 3.1127 for the LFPW and
2.8801 and 2.0417 for the MORPH, respectively. The proposed model is proven to have
shown the best results, and thus compensating shape and texture variations. The
information obtained from these model yield semantic criteria to be used in content
based image retrieval, further improvement can be made to make the proposed model
even more robust and applicable.
|
|---|