IRJET - Real Time Facial Analysis using Tensorflowand OpenCV
0 likes29 views
This document presents a real-time facial analysis system using TensorFlow and OpenCV. The system can detect facial expressions, age, and gender from images and video in real-time. It uses deep learning models trained on facial datasets to analyze faces. The system is designed for applications like security, attendance tracking, and finding lost children. It works by extracting facial features from images, applying preprocessing techniques, classifying faces, and making predictions about attributes. The document discusses the methodology, existing techniques like PCA and HMM, the proposed system architecture, sample code, and conclusions.
![International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3083
REAL TIME FACIAL ANALYSIS USING TENSORFLOWAND OPENCV
Lokesh S1, Nithish singh A2, Raja D3, Janani S4
1,2,3Computer science and Engineering , Adhiyamaan college of engineering , Tamil Nadu , India
4Assistant Professor, Department of Computer Science, Adhiyamaan college of engineering, Tamil Nadu, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - The main objective of this project to develop a
real time facial analysis using opencv and tensorflow for
artisum people where user can see the facial expression, age
and gender of the male and female. Face regonition and
analysis from an image and video. Most signal or image
processing algorithms should be designed with real time
execution in mind.Facial features are considered as one of
the important personal characteristics. This can be used in
many applications, such as face recognition and age
estimation. The value of these applications depends in
several areas, such as security applications, law
enforcement applications, and attendance systems. In
addition, facial features are particularly the key usage in the
finding of lost child. Present applications have achieved a
high level of accuracy. This paper provides a survey of face
recognition, including the age estimation, which was
discussed. Moreover, the research outlines several
challenges faced in face recognition area that had been
explored. The research also provides a landscape mapping
based on integrating into a critical and coherent taxonomy.
In the methodology sections, the exploration the
accomplished via a deep focused in every single article in
‘‘Face Recognition’’, then ‘‘Age Estimation’’, and later in
‘‘Facial Features’’. The ‘‘Articles extraction’’ is mining from
diverse sources, such as Web of Science, ACM, IEEE, Science
Direct, and Springer databases. The research covers overall
72 articles; 32/72 articles were face recognition. Moreover,
39/72 of the articles were for age estimation. A comparison
based on the objectives of the approaches is presented to
underline the taxonomy. Ending by research conclusion on
face techniques contributes to the understanding of the
recognition approaches, which can be used in future
researches. The research concluded that face techniques’
performance is distinct from one data set to another. This
paper contributes to display gaps for other researchers to
join this line of research.
Key Words: Face recognition, age estimation, aging and
facial expression
1. INTRODUCTION
There are few facial features used in facial technology.
This technology is used to recognize the exterior organs of
the body such as mouth, eyes and also gender of a person.
The major source of data that is used in this technology is
the features extracted from the organs. There are three
levels of features used. The foremost factor is to extract in
depth micro level features that includes moles on body,
birth marks [4]. This technology is widely used in various
domain such as stage estimation etc. In most of the models
the major process to be considered is as the stage of the
person changes from time to time accordingly the
appearance of a person change. The major flaw in this
technique is wrinkles formed on face as the person grows
older. The technology is evolved over a time to specify the
age of a person based on the visual appearance even
though the birth date of a person is not known. Anyone
can calculate the age of a person by knowing the birth
date, but this technology is specifically designed to
estimate the age of a person based on the facial looks that
too mainly if a person face is with wrinkles then there
some advantage and disadvantage involved in it [6]. Aging
is a common factor that is affecting and irrepressible
process in each and every individual life throughout the
life span [7]. This type of technology is also used in
criminal activities to find out lost children. There are
already few major challenges that were enforced. This
paper provides clear cut details on the architecture,
advantages and limitations of facial recognition technique.
There are n number of domains using this type of
recognition such as in biology, security purpose and so on.
The same technology is also used by human forensics to
recognize the late person by collaborating the tissue in
face [8]. As the list prolong in the usage of this technique,
the one such major application uses this method is in
security purpose such as biometric identification [9]. The
collection of named data is one such problem faced in this
methodology is more complex compared to age detection
based on person looks [10], [11]. This paper explains
about the disadvantages faced in existing system and how
it can be overcome by using proposed system. This paper
also provide an edge to the challenges that are
experienced in facial recognition, architecture used.
1.1 METHODOLOGY
There are two terminology that are most preferably used
in this method such as face and age estimation. This also
includes features extraction from face to determine the
age of the person that is caused because of ageing scenario
that every person has to undergo in his/her lifespan
INFORMATION SOURCES
Even though there are so many existing database were
found to get an immense knowledge on the problems that
are encountered in the field of facial recognition
technology and we have referred to few articles such a
Springer, IEEE Xplore, Science Direct database, WoS
service, ACM DL and list prolongs.](https://image.slidesharecdn.com/irjet-v7i2663-201130062646/85/IRJET-Real-Time-Facial-Analysis-using-Tensorflowand-OpenCV-1-320.jpg)
![International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3084
A. SELECT PAPER
The selection of the paper to get an idea of how existing
system work depends two modes: iteration and filtering.
During iteration the unwanted and irrelevant paper are
removed/hidden similar to the technical word abstraction
in technical term. Filtering phase is completely following
this stage.
B. ELIGIBILITY CRITERIA
To develop the proposed system as a developers the
selection of papers and papers referred are carried from
2010 to 2018.
2. EXISTING SYSTEMS
LINEAR DISCRIMINATE ANALYSIS
LDA is used to minimize features that will be
overwhelming and mostly not used in this technique. The
resultant obtained is linear classifier. To detect the facial
features properly, this analysis requires more number of
pixels.
PRINCIPAL COMPONENT ANALYSIS
PCA is a procedure that converts correlated variables into
uncorrelated variables. This uncorrelated variables are
known as principal components. The first uncorrelated
components says about the variation in the data and
subsequent components says about other variations that
can be expected. PCA is widely used to develop predictive
models and so on. With the help of PCA the Eigen value
decomposition is performed. By skimpy the internal
structure of the data the variance present in it is known.
HIDDEN MARKOV MODEL
HMM is a statistical model that describes the event that
are not directly observable, but the events are depending
interior factors. The observed event is called as a ‘symbol’
and the factor underlying the observation is a ‘state’. This
type of model is preferably used in bioinformatics,
biological sequence analysis.
3. PROPOSED SYSTEM AND TECHNIQUE
System accuracy is influenced by different factors in that
one such pattern is number of pixels combined to a form
an image. Before supplying any image to face recognition
system the foremost attention should be taken by applying
all pre-processing techniques. Depending upon how the
algorithm is trained the face recognition system works.
Consider if an algorithm is trained to detect a person in a
bright room then it is impossible for a person to be
detected in a dark room. There are various other kinds of
problems be faced in this method such as all the pixel
coordinate should be same. Hence a proper preprocessing
filter is recommended. The proposed system uses Eigen
values using grayscale images. Histogram is applied after
conversion of image to grayscale is done. The more of
preprocessing techniques applied, the more the clarity of
image is found and thus results obtained is also less error
prone. The image that is sent via file or video is classified
to detect the faces. The classification is done faster and
with a good efficiency.
4. SAMPLE CODE
A. Training Set
Def assure_path_exists(path):
dir = os.path.dirname(path)
if not os.path.exists(dir):
os.makedirs(dir)
recognizer = cv2.face.LBPHFaceRecognizer_create()
Detector=cv2.CascadeClassifier("haarcascade_frontalface_
de fault.xml");
def getImagesAndLabels(path):
imagePaths = [os.path.join(path,f)
for f in os.listdir(path)]
faceSamples=[]
ids = []
for imagePath in imagePaths:
PIL_img = Image.open(imagePath).convert('L')
img_numpy = np.array(PIL_img,'uint8')
id = int(os.path.split(imagePath)[-1].split(".")[1])
faces = detector.detectMultiScale(img_numpy)
for (x,y,w,h) in faces:
faceSamples.append(img_numpy[y:y+h,x:x+w])
ids.append(id)
return
faceSamples,ids faces,ids = getImagesAndLabels('dataset')
recognizer.train(faces, np.array(ids))
assure_path_exists('trainer/')
recognizer.save('trainer/trainer.yml')
B. Face Datasets
def assure_path_exists(path):
dir = os.path.dirname(path)
if not os.path.exists(dir):](https://image.slidesharecdn.com/irjet-v7i2663-201130062646/85/IRJET-Real-Time-Facial-Analysis-using-Tensorflowand-OpenCV-2-320.jpg)
![International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3085
os.makedirs(dir)
vid_cam = cv2.VideoCapture(0)
face_detector=cv2.CascadeClassifier('haarcascade_frontalf
ace_default.xm l')
face_id = 1 count = 0
assure_path_exists("dataset/")
while(True): _,
image_frame = vid_cam.read()
gray=cv2.cvtColor(image_frame, cv2.COLOR_BGR2GRAY)
faces = face_detector.detectMultiScale(gray, 1.3, 5)
for (x,y,w,h) in faces: cv2.rectangle(image_frame, (x,y),
(x+w,y+h), (255,0,0), 2)
count += 1 cv2.imwrite("dataset/User." + str(face_id) +
str(count) + ".jpg", gray[y:y+h,x:x+w])
cv2.imshow('frame', image_frame)
if cv2.waitKey(100) & 0xFF == ord('q'):
break
elif
count>100:
break
vid_cam.release()
cv2.destroyAllWindows()
C. Face Detection
def assure_path_exists(path):
dir = os.path.dirname(path)
if not os.path.exists(dir):
os.makedirs(dir)
recognizer = cv2.face.LBPHFaceRecognizer_create()
assure_path_exists("trainer/")
recognizer.read('trainer/trainer.yml')
cascadePath = "haarcascade_frontalface_default.xml"
faceCascade = cv2.CascadeClassifier(cascadePath);
font = cv2.FONT_HERSHEY_SIMPLEX
cam = cv2.VideoCapture(0)
while True:ret,
im=cam.read()
gray = cv2.cvtColor(im,cv2.COLOR_BGR2GRAY)
faces = faceCascade.detectMultiScale(gray, 1.2,5)
for(x,y,w,h) in faces: cv2.rectangle(im, (x-20,y-20),
(x+w+20,y+h+20), (0,255,0), 4) Id, confidence =
recognizer.predict(gray[y:y+h,x:x+w])
if(Id == 1):
Id = "ARJUN {0:.2f}%".format(round(100 - confidence, 2))
cv2.rectangle(im, (x-22,y-90), (x+w+22, y-22), (0,255,0),
1) cv2.putText(im, str(Id), (x,y-40), font, 1, (255,255,255),
3) cv2.imshow('im',im)
if cv2.waitKey(10) & 0xFF == ord('q'): break
cam.release() cv2.destroyAllWindows()
5. CONCLUSION
In this paper, by using opencv as a set of library
programming functions a system is developed to estimate
the age of a person by extracting the features in face. A
dataset that contains a set of images are defined and
trained before recognizing process proceeds. Haar cascade
algorithm is used for detection. In upcoming generations
few more features are included to enhance the changes for
a proper recognition by overcoming the problems such as
wrinkles on a face etc.
REFERENCES
[1]https://www.youtube.com/watch?list=PLQVvvaa0QuD
f KTOs3Keq_kaG2P55YRn5v&v=OGxgnH8y2NM
[2] https://opencv.org/
[3]https://docs.opencv.org/3.4/d7/d8b/tutorial_py_face_
d etection.html
[4]https://pythonprogramming.net/haar-cascade-
objectdetection-python-opencv-tutorial/
[5] .Open Source Computer Vision Library Reference
Manual-intel [Media]
[6] Tej Pal Singh, “Face Recognition by using Feed
Forward Back Propagation Neural Network”, International
Journal of Innovative Research in Technology & Science,
vol.1, no.1
[7] N.Revathy, T.Guhan, “Face recognition system using
back propagation artificial neural networks”, International
Journal of Advanced Engineering Technology, vol.3, no. 1,
2012.
[8] M.A.Turk and A.P. Pentaland, “Face Recognition Using
Eigenfaces”, IEEE conf. on Computer Vision and Pattern
Recognition, pp. 586-591, 1991](https://image.slidesharecdn.com/irjet-v7i2663-201130062646/85/IRJET-Real-Time-Facial-Analysis-using-Tensorflowand-OpenCV-3-320.jpg)
IRJET - Real Time Facial Analysis using Tensorflowand OpenCV
- 1. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072 © 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3083 REAL TIME FACIAL ANALYSIS USING TENSORFLOWAND OPENCV Lokesh S1, Nithish singh A2, Raja D3, Janani S4 1,2,3Computer science and Engineering , Adhiyamaan college of engineering , Tamil Nadu , India 4Assistant Professor, Department of Computer Science, Adhiyamaan college of engineering, Tamil Nadu, India ---------------------------------------------------------------------***---------------------------------------------------------------------- Abstract - The main objective of this project to develop a real time facial analysis using opencv and tensorflow for artisum people where user can see the facial expression, age and gender of the male and female. Face regonition and analysis from an image and video. Most signal or image processing algorithms should be designed with real time execution in mind.Facial features are considered as one of the important personal characteristics. This can be used in many applications, such as face recognition and age estimation. The value of these applications depends in several areas, such as security applications, law enforcement applications, and attendance systems. In addition, facial features are particularly the key usage in the finding of lost child. Present applications have achieved a high level of accuracy. This paper provides a survey of face recognition, including the age estimation, which was discussed. Moreover, the research outlines several challenges faced in face recognition area that had been explored. The research also provides a landscape mapping based on integrating into a critical and coherent taxonomy. In the methodology sections, the exploration the accomplished via a deep focused in every single article in ‘‘Face Recognition’’, then ‘‘Age Estimation’’, and later in ‘‘Facial Features’’. The ‘‘Articles extraction’’ is mining from diverse sources, such as Web of Science, ACM, IEEE, Science Direct, and Springer databases. The research covers overall 72 articles; 32/72 articles were face recognition. Moreover, 39/72 of the articles were for age estimation. A comparison based on the objectives of the approaches is presented to underline the taxonomy. Ending by research conclusion on face techniques contributes to the understanding of the recognition approaches, which can be used in future researches. The research concluded that face techniques’ performance is distinct from one data set to another. This paper contributes to display gaps for other researchers to join this line of research. Key Words: Face recognition, age estimation, aging and facial expression 1. INTRODUCTION There are few facial features used in facial technology. This technology is used to recognize the exterior organs of the body such as mouth, eyes and also gender of a person. The major source of data that is used in this technology is the features extracted from the organs. There are three levels of features used. The foremost factor is to extract in depth micro level features that includes moles on body, birth marks [4]. This technology is widely used in various domain such as stage estimation etc. In most of the models the major process to be considered is as the stage of the person changes from time to time accordingly the appearance of a person change. The major flaw in this technique is wrinkles formed on face as the person grows older. The technology is evolved over a time to specify the age of a person based on the visual appearance even though the birth date of a person is not known. Anyone can calculate the age of a person by knowing the birth date, but this technology is specifically designed to estimate the age of a person based on the facial looks that too mainly if a person face is with wrinkles then there some advantage and disadvantage involved in it [6]. Aging is a common factor that is affecting and irrepressible process in each and every individual life throughout the life span [7]. This type of technology is also used in criminal activities to find out lost children. There are already few major challenges that were enforced. This paper provides clear cut details on the architecture, advantages and limitations of facial recognition technique. There are n number of domains using this type of recognition such as in biology, security purpose and so on. The same technology is also used by human forensics to recognize the late person by collaborating the tissue in face [8]. As the list prolong in the usage of this technique, the one such major application uses this method is in security purpose such as biometric identification [9]. The collection of named data is one such problem faced in this methodology is more complex compared to age detection based on person looks [10], [11]. This paper explains about the disadvantages faced in existing system and how it can be overcome by using proposed system. This paper also provide an edge to the challenges that are experienced in facial recognition, architecture used. 1.1 METHODOLOGY There are two terminology that are most preferably used in this method such as face and age estimation. This also includes features extraction from face to determine the age of the person that is caused because of ageing scenario that every person has to undergo in his/her lifespan INFORMATION SOURCES Even though there are so many existing database were found to get an immense knowledge on the problems that are encountered in the field of facial recognition technology and we have referred to few articles such a Springer, IEEE Xplore, Science Direct database, WoS service, ACM DL and list prolongs.
- 2. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072 © 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3084 A. SELECT PAPER The selection of the paper to get an idea of how existing system work depends two modes: iteration and filtering. During iteration the unwanted and irrelevant paper are removed/hidden similar to the technical word abstraction in technical term. Filtering phase is completely following this stage. B. ELIGIBILITY CRITERIA To develop the proposed system as a developers the selection of papers and papers referred are carried from 2010 to 2018. 2. EXISTING SYSTEMS LINEAR DISCRIMINATE ANALYSIS LDA is used to minimize features that will be overwhelming and mostly not used in this technique. The resultant obtained is linear classifier. To detect the facial features properly, this analysis requires more number of pixels. PRINCIPAL COMPONENT ANALYSIS PCA is a procedure that converts correlated variables into uncorrelated variables. This uncorrelated variables are known as principal components. The first uncorrelated components says about the variation in the data and subsequent components says about other variations that can be expected. PCA is widely used to develop predictive models and so on. With the help of PCA the Eigen value decomposition is performed. By skimpy the internal structure of the data the variance present in it is known. HIDDEN MARKOV MODEL HMM is a statistical model that describes the event that are not directly observable, but the events are depending interior factors. The observed event is called as a ‘symbol’ and the factor underlying the observation is a ‘state’. This type of model is preferably used in bioinformatics, biological sequence analysis. 3. PROPOSED SYSTEM AND TECHNIQUE System accuracy is influenced by different factors in that one such pattern is number of pixels combined to a form an image. Before supplying any image to face recognition system the foremost attention should be taken by applying all pre-processing techniques. Depending upon how the algorithm is trained the face recognition system works. Consider if an algorithm is trained to detect a person in a bright room then it is impossible for a person to be detected in a dark room. There are various other kinds of problems be faced in this method such as all the pixel coordinate should be same. Hence a proper preprocessing filter is recommended. The proposed system uses Eigen values using grayscale images. Histogram is applied after conversion of image to grayscale is done. The more of preprocessing techniques applied, the more the clarity of image is found and thus results obtained is also less error prone. The image that is sent via file or video is classified to detect the faces. The classification is done faster and with a good efficiency. 4. SAMPLE CODE A. Training Set Def assure_path_exists(path): dir = os.path.dirname(path) if not os.path.exists(dir): os.makedirs(dir) recognizer = cv2.face.LBPHFaceRecognizer_create() Detector=cv2.CascadeClassifier("haarcascade_frontalface_ de fault.xml"); def getImagesAndLabels(path): imagePaths = [os.path.join(path,f) for f in os.listdir(path)] faceSamples=[] ids = [] for imagePath in imagePaths: PIL_img = Image.open(imagePath).convert('L') img_numpy = np.array(PIL_img,'uint8') id = int(os.path.split(imagePath)[-1].split(".")[1]) faces = detector.detectMultiScale(img_numpy) for (x,y,w,h) in faces: faceSamples.append(img_numpy[y:y+h,x:x+w]) ids.append(id) return faceSamples,ids faces,ids = getImagesAndLabels('dataset') recognizer.train(faces, np.array(ids)) assure_path_exists('trainer/') recognizer.save('trainer/trainer.yml') B. Face Datasets def assure_path_exists(path): dir = os.path.dirname(path) if not os.path.exists(dir):
- 3. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072 © 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3085 os.makedirs(dir) vid_cam = cv2.VideoCapture(0) face_detector=cv2.CascadeClassifier('haarcascade_frontalf ace_default.xm l') face_id = 1 count = 0 assure_path_exists("dataset/") while(True): _, image_frame = vid_cam.read() gray=cv2.cvtColor(image_frame, cv2.COLOR_BGR2GRAY) faces = face_detector.detectMultiScale(gray, 1.3, 5) for (x,y,w,h) in faces: cv2.rectangle(image_frame, (x,y), (x+w,y+h), (255,0,0), 2) count += 1 cv2.imwrite("dataset/User." + str(face_id) + str(count) + ".jpg", gray[y:y+h,x:x+w]) cv2.imshow('frame', image_frame) if cv2.waitKey(100) & 0xFF == ord('q'): break elif count>100: break vid_cam.release() cv2.destroyAllWindows() C. Face Detection def assure_path_exists(path): dir = os.path.dirname(path) if not os.path.exists(dir): os.makedirs(dir) recognizer = cv2.face.LBPHFaceRecognizer_create() assure_path_exists("trainer/") recognizer.read('trainer/trainer.yml') cascadePath = "haarcascade_frontalface_default.xml" faceCascade = cv2.CascadeClassifier(cascadePath); font = cv2.FONT_HERSHEY_SIMPLEX cam = cv2.VideoCapture(0) while True:ret, im=cam.read() gray = cv2.cvtColor(im,cv2.COLOR_BGR2GRAY) faces = faceCascade.detectMultiScale(gray, 1.2,5) for(x,y,w,h) in faces: cv2.rectangle(im, (x-20,y-20), (x+w+20,y+h+20), (0,255,0), 4) Id, confidence = recognizer.predict(gray[y:y+h,x:x+w]) if(Id == 1): Id = "ARJUN {0:.2f}%".format(round(100 - confidence, 2)) cv2.rectangle(im, (x-22,y-90), (x+w+22, y-22), (0,255,0), 1) cv2.putText(im, str(Id), (x,y-40), font, 1, (255,255,255), 3) cv2.imshow('im',im) if cv2.waitKey(10) & 0xFF == ord('q'): break cam.release() cv2.destroyAllWindows() 5. CONCLUSION In this paper, by using opencv as a set of library programming functions a system is developed to estimate the age of a person by extracting the features in face. A dataset that contains a set of images are defined and trained before recognizing process proceeds. Haar cascade algorithm is used for detection. In upcoming generations few more features are included to enhance the changes for a proper recognition by overcoming the problems such as wrinkles on a face etc. REFERENCES [1]https://www.youtube.com/watch?list=PLQVvvaa0QuD f KTOs3Keq_kaG2P55YRn5v&v=OGxgnH8y2NM [2] https://opencv.org/ [3]https://docs.opencv.org/3.4/d7/d8b/tutorial_py_face_ d etection.html [4]https://pythonprogramming.net/haar-cascade- objectdetection-python-opencv-tutorial/ [5] .Open Source Computer Vision Library Reference Manual-intel [Media] [6] Tej Pal Singh, “Face Recognition by using Feed Forward Back Propagation Neural Network”, International Journal of Innovative Research in Technology & Science, vol.1, no.1 [7] N.Revathy, T.Guhan, “Face recognition system using back propagation artificial neural networks”, International Journal of Advanced Engineering Technology, vol.3, no. 1, 2012. [8] M.A.Turk and A.P. Pentaland, “Face Recognition Using Eigenfaces”, IEEE conf. on Computer Vision and Pattern Recognition, pp. 586-591, 1991