Research On Real Time Facial Feature Point Detection On Mobile Device

Facial feature point detection is a fundamental issue in image processing,bio-metric authentication and augment reality.Although some models for facial point detection achieve very good result,they suffer from poor speed when it comes to a mobile device like smart phone.Smart phones can not use a Central Processing Unit like Personal Computer considering the battery capacity and other limitations.This thesis aims to explore more accurate and fast facial feature point detection algorithms on mobile devices.The main works are as follows:1.Before fitting the Constrained Local Model to a image containing a face,we first use a skin detector together with an Adaboost Face Detector to reduce the search area.2.In the shape training step for Constrained Local Model,we use PCA as a dimensional reduction method.In the experiment,we found that after transformation,scaling and rotation step,there are still some outliers.We propose a method to detect the outliers and delete them for a better performance as PCA is sensitive to outliers.3.The most time consuming aspect in Constrained Local Model is the recursive step for finding the best facial point location.In the experiment,we have to iterate more than 20 times to achieve a relative good result and this is unacceptable in a real time application.We propose a method that can avoid the time consuming aspect and achieve better performance.4.We implement the proposed algorithm in JavaScript which allows our program runs in a mobile platform easily.According theoretical analysis and experiment results,the modified constrained local model improves the location accuracy and fit speed.It achieves a real-time performance on mobile devices.