Using machine learning algorithms to detect smile in pictures based on geometric features

العناوين الأخرى

استخدام التعلم الألي لرصد الابتسامة اعتمادا على الخصائص الهندسية لشفاه الفم

مقدم أطروحة جامعية

Muhammad, Dua Ahmad Jibril

مشرف أطروحة جامعية

Nassar, Muhammad Uthman

الجامعة

جامعة عمان العربية

الكلية

كلية العلوم الحاسوبية و المعلوماتية

القسم الأكاديمي

قسم علم الحاسوب

دولة الجامعة

الأردن

الدرجة العلمية

ماجستير

تاريخ الدرجة العلمية

2018

الملخص الإنجليزي

Usage of Digital images applications have become one of the most used applications these days.

It is increased rapidly and developers are working continuously to improve them by adding new features to make them easier to use and more user friendly.

One of the most required features in those applications is smile detection.

Searching literature and researches about digital smile detection revealed a lot of methods.

However, none of these applies angles based approach to detect smile in images using machine learning software.

This research is devoted to detect smile in pictures passing through stepwise stages.

first stage use Flandmark algorithm used to detect human faces in pictures, followed by detecting human lips by using the same algorithm.

In the second stage, lips landmarks are determined and used to draw lines and generate five angles for the lips.

Those angles are used create new parameter dataset that’s will be the first of its kind.

The third stage has three separated levels; according to the number of angles to be used, level one uses two angles (horizontal angles) with/without the distance between points.

In the second level, three angles were used also with/without the distance between points, and finally we will work with five angles in the third level.

At the third stage, we create the new dataset from the previous angles by changing it to parameters form with (.

arff) file type.

At the fourth stage, we divide our method into many probabilities to determine the best probability to detect the smile; these probabilities depend on the number of parameter that used in each experiment: Using ang1_ang2 and the two distances, using ang1_ang3 and the two distances, using ang2_ang4 and the two distances, using ang3_ang4_ang5 and the distances, using all angles and the distances, and finally; Using all angles without distances.

all probabilities are tested three times, the first and the second will be tested by cross validation method using many Weka classifiers with modifications and without modifications on classifiers properties, And the third one by using training/testing method with properties modifications.In this research, nine classifiers were used: Bayes Net, Naïve Bayes, SMO, IBK, Decision Table, BF Tree, J48, Random Tree, and Rep Tree.

The number of experiments reached to 594 to include all factors that can affect the results and accuracy of the research.

The results that we get indicates that our proposed method is more efficient than 2DPCA, Adaboost, PCA and Shan et al., and give us 100% accuracy when using IBK and Random tree classifiers with training/testing method with any combination of angles.

Also we found that SMO classifier in most cases can detect all non-smiley faces in the dataset so it’s the best classifier to be used to detect non-smile faces but cannot detect any smiley face.

التخصصات الرئيسية

تكنولوجيا المعلومات وعلم الحاسوب

الموضوعات

• التعلم الآلي
• الخوارزميات
• الصور الرقمية

عدد الصفحات

77

قائمة المحتويات

• APA
• MLA
• AMA

لغة النص

الإنجليزية

نوع البيانات

رسائل جامعية

رقم السجل

BIM-932344