簡介 - 國立台北大學

advertisement
國立台北大學資訊工程學系專題報告
專題題目: Using Local Binary Pattern &SVM to achieve
Face gender identification
專題組員:曾奕傑 包黛安
指導老師:楊致芳老師
1. 簡介
What Face Recognition:
 Face recognition is a method used to classify data based on
face information identify or verify a person using selected
features from the image and a facial database.
 Data could be used for gender identification, age or face detection.
 In our experiment , we are used the features to identify the
person’s gender.
 Real life practical uses of face recognition: Criminal database
search, Security safe door etc.
Motivation:
 To figure out the best procedure for implementing gender
identification using the SVM.
 To determine the effectiveness of the SVM when performing gender
identification.
 To understand the implementation of Local Binary Pattern and the
SVM.
 To carryout investigations on previous gender identification
results.
Experiment, equipment and platform:
 Application software: LIBSVM
 Database: colour Feret
 Hardware: personal laptops and computers in the lab
 Running software: Matlab
2. 專題進行方式
藉著將人臉圖片 LPB 化後用不同的 vector 形式傳入 svm 以得到不同的分辨
結果來分析何種方式較好何種較差,為什麼?如何改進,是否有錯誤產生?
Experimental procedure:
LBP Algorithm r 介紹:
LBP transform example:
→
LIBSVM 介紹:
 LIBSVM is an integrated software for support vector classification,
regression and distribution estimation. It supports multi-class
classification.
 Developed by the National Taiwan University Professor Lin
Zhiren(林智仁), the principle is the Support Vector Machine (SVM)
SVM code:
 model = svmtrain(label, feature vector);
Training model
 [predict_label, accuracy, dec_values] = svmpredict(label, feature
vector, model);
Predict data
以下是原先設想的 vector 型態分成三種:
Original Feature Vectors:
 Link (method 1) – With this method, each picture is separated into
25 equal bins. The histogram (containing the local binary pattern)
for each bin is calculated. The histograms are then linked to each
other and sent to the SVM.
 Threshold (method 2) – With this method, the bins are added
together, these added bins are then averaged. The bins that were
combined to determine the average histogram are then compared to
said average histogram. This results in a 26 binary digit which
is calculated based on whether or not they are greater or less than
the average histogram.
 Link and Threshold (method 3) – This method comprises of a
combination of both the link and threshold method.
Experiments carried out on the original feature vectors
這是我們對先前設想的部分做改進的方式
 In our experiment, we categorised the data. This was done by
comparing the data to the threshold and then placing it in a category
that fell either below or above the threshold. It’s
classification depended on certain boundaries that were set when
compared to the threshold multiplied by different variables. Our
experiment was done on method 2 (threshold) and method 3(link
combined with threshold).
3. 主要成果與評估
Original Results:
Expected Results:
 Prior to carrying out our experiments, we felt that both methods
would decrease in their ability to identify the gender as the number
of sections increased.
Results after Experiment:
Observations and Explanations
Q1: Regardless of the amount of times the data in method 2 was partitioned,
the test results were basically the same.
A1: Method 2’s data is too rough. It appears that the data that was sent
to the SVM didn’t contain enough information in order to yield a better
outcome. Partitioning the data had no great impact on the outcome. The
reason for this is that regardless of the amount of partitions used, the
data sent was still unrefined and contained just binary digits.
Q2: When the data was placed into four categories using the third method
(link + threshold), there was an increase in the ability to identify the
gender using the cut images (84.1424%). However, this resulted in a
decrease in the ability to identify the gender using the uncut images.
(79.59594%)
A2: An increase in the categories means there is also an increase in the
information that was sent to the SVM. The cut images contains noise, the
added information complicates the problem which made the task more
difficult. On the other hand, the added information proved to be more
beneficial to the cut images since the absence of noise meant that the
information sent was not too much
Q3: When the data was placed into categories of 8, there was a decrease
in the ability to identify the gender in both the cut (55.57144%) and uncut
(60.101%)images. However, the uncut images yielded a better result than
the cut images.
A3: As noted earlier, an increase in the categories also results in an
increase in the information that is being sent to the SVM. When looking
at the cut images, we realised that we encountered an overfitting problem.
The information was increased to the extent that the parameters that
differentiated the males and females became so close that it was difficult
to distinguish between the two.
A: The uncut images resulted in a better identification percentage than
the cut images because the presence of noise helped to eliminate the
overfitting problem. Examples of noise are shoulders and hair.
Uncut
Cut
Conclusion:
 Senior’s best result:
Train:99.32144%
Test:82.7143%
 Our best result:
Train:100%
Test:84.14284%
 Method 3 proves to be better than both the link and the threshold
method. With the link method, one picture has 650 feature vectors.
On the other hand the threshold method has 26 binary feature vectors.
The data relation using the threshold method is enhanced, thus it’s
more efficient than the link method in identifying the gender.
 In order for method 3 to be more effective, that is, if we want to
see an increase in its accuracy in correctly identifying the gender,
we should increase the size of the database.
Further Experiment:
 We propose that a combination of the link and threshold method
(method 3) be used with a larger database that has at least 16
categories.
4. 結語與展望
這次的專題製作讓我們學到了很多寶貴的經驗,雖然我們的實驗嚴格說起來才剛
剛邁入真的要進行實驗分析與改進的部分,這兩個學期裡從一開始的全然無知道
現在的可以真的進行一些稱得上是實驗的過程其實要感謝學長以即老師的鼓勵,
如論做的好壞,老師都會給予正面的想法以及提出我們兩個所不足的地方,雖然
結果不盡完美,但是終究是個經驗,而跨文化的合作也讓我們兩個產生許多不同
的體驗,希望未來有機會有人會願意完善這個實驗。
5. 銘謝
We would like to thank Professor Yang, the seniors in the lab and our
friends and family for their unwavering support and well wishes.
6. 參考文獻

Ahomen, T
& Hadid, A & Pietikainen, M. 2006. Face Description with
Binary Patterns: Application to face recognition.

Local
IEEE 1 - 13
Ojala, T & Pietkainen, M & Maepaa, T (2002, July) IEEE TRANSACTIONS ON
PATTERN RECOGNITION AND MACHINE INTELLIGENCE: Multiresolution and
Rotation Invariant Texture Classification with Local Binary Patterns,
),
971 – 976, Retrieved from:
http://www.rafbis.it/biplab15/images/stories/docenti/Danielriccio/Articol
iriferimento/LBP.pdf

Hui-cheng Lian & Bao-Liang Lu (2006) ADVANCES IN NEURAL NETWORKS –
ISNN
2002: Multi-view Gender Classification using Local Binary Patterns
and
Support Vector Machines, 3972(2006) – DOI: 10.1007/11760023_30

"SVM - Support Vector Machines." SVM - Support Vector Machines. Web. 28 May
2012. <http://www.dtreg.com/svm.htm>.
Download