Face retrieval defined by Zhong Wu et al., as
“Given images or frames of video as query, our goal is to retrieve video
containing faces of the same person appearing in the query video from face
video database”. The foremost intention is to search and retrieve human face in
video from given query of face frames (video). The various existing approaches
and techniques have been used more than one parameter along with face such as
voice, gait, tag, text, content, etc for various applications.
The various problems are associated with the existing
system such as, appropriate real time face detection and tracking algorithm, feasible
feature extraction technique with optimum features, fast and accurate face
matching (recognition/ retrieval) technique. Whereas, the various challenges
are involved such as illumination, facial expression, occlusion, background,
presence and absence of face, etc. in video database.
To overcome the problems in existing techniques, the
proposed research work has implemented and experiments performed on human face
retrieval system which has capable to detect and train face from video. The proposed
technique used minimum features for feature extraction based on discrete
Wavelet Transformation. The performance of proposed technique is calculated on
various parameters and compared with well known methods/techniques on same
The two algorithms have been implemented which involved
steps like face detection-tracking, pre-processing and matching (recognition/
retrieval). The algorithms based on Haar Wavelet for feature extraction using Leonardo
de Vinci technique, which extract eyes, mouth and nose portion of the face. The
face is detected on every frame of video using Viola and Jones real time face
detection algorithm. The algorithm is designed and implemented for
face matching (recognition/ retrieval) using minimum difference between 24
features of trained and test datasets.
The standard video databases UO Face video database is used for experimental
purpose. These database includes conditions such as low resolution,
motion blur, out-of focus factor, facial expression variation,
facial orientation variation, occlusions, under real indoor and outdoor
illumination conditions, head rotations, partial occlusion, face partly leaving
the field of view, and large scale changes, etc. UO Face video database has
contained 12 videos. OU face video is tested for 240 times for proposed
algorithm. The similar tests have been performed for KPCA, KPCA+, LDA and LDA+.
To allow fair comparison, in addition to accuracy,
various parameters are calculated such as error rate, evaluation or retrieval
time for the algorithm. The proposed technique has provided mean accuracy rate
of 83.52%, on OU face video database. The mean accuracy rate and mean error
rate of proposed technique has shown better results as compared to KPCA, KPCA+,
LDA and LDA +. Precisely, the mean evaluation time (retrieval time) on equal
parameter has 2.12sec on OU face video database, which is 12.57sec on similar
database using KPCA+.
The key requirement of face retrieval is number of
features using feature extraction technique, the basic problem of intensity
based algorithm is the selection of features. The accuracy of the system is
directly depend on number of features and inversely proportional to complexity.
The proposed work has overcome the problem of feature selection using discrete
More precisely, it is observed that proposed technique
using discrete wavelet transformation required less evaluation time and space
than KPCA and LDA and more accurate results shown (Mean accuracy rate and Mean
error rate) on video databases.
Real time face retrieval is very powerful and useful
system in various applications such as security and surveillance system, state
of art, character searching or browsing from video, video annotation, video
retrieval, video summarization, people identification, facial expression,
gender identification and various other applications. The technique will be
used for online applications like tagging, web based character or object
browsing, criminal identification etc.
In future this
research work will be extended to improve the results in terms of mean accuracy
and mean error rate. This research work will also be extended for multiple
persons per video, real videos and high resolution video data for online and
The research will never stop, it always searches for
improvement. The proposed research work will be elaborated with other
significant features of the human using biometrics.