A Face Morph Detection Method Based On Neural Networks And Occlusion Test

Face morphing is a very severe threat to Automatic Border Control(ABC)systems.Such systems are getting world-wide application and recognition at sea and airports to allow migrants to access or leave nations or states.ABCs or eGates are slowly replacing personnel tasked with checking the validity of immigrants’ information but could be beaten by face morphing.Therefore,this research focuses on the detection of these face morphs so as to protect countries from frauds or people that could possess malicious intentions while preserving the integrity of these systems that are so expensive and the main research contents are as follows:First,we explain what a face morph is and how it can be created,showcasing a number of channels used to obtain face morphs.We then create and state our datasets used in this research whereby one consists of 5 subjects and was manually generated.The subjects chosen are popular people of all genders and races.The other dataset was obtained over the internet and it has about 140,000 images both genuine and morphed,100,000 of which are for the training dataset.All images were scaled down to 224 X 224 pixels and converted into gray scale.Secondly,we define and explain our proposed methods used in this research in detail that is occlusion detection and neural networks.For occlusion detection,we implemented the FSG-FD method which combines multi-scale features,context and feature enhancement(SG-net,a feature enhancement feature for FSG-FD)to detect occlusions.For neural networks,we implemented the VGG19 architecture which consists of 16 convolutional layers,3 fully connected layers,5 max pool layers and 1soft max layer to perform its classification tasks.For this research we pre-trained our system and increased the convolutional layers of VGG19 by an additional 3 convolutional layers to increase its classification capabilities.Take note that each of the proposed methods can detect face morphs to some extent.In this study,we combine the two and simultaneously apply them to detect face morphs,which further improves the detection accuracy.Finally,the methods applied gave better results with FSG-FD showing improvement amongst other occlusion detection methods like MTCNN and RCNN methods used before.The same is true for the VGG19 architecture deployed for our neural network branch.Tests and comparisons with other findings in relatable fields are shown in a series of tables and figures in chapter 4.The failures or short comings of one method seem to be neglected or negated by the other method because of the combination of the methods used in this research.For example the FSG-FD method can ignore face features like cheeks,chin and forehead and only focus on features like mouth,nose and mouth in serious occlusions but the pre-trained VGG19 architecture considers the entire image.Our VGG19 method achieved about 6.7%accuracy improvement than VGG16 architecture and MTCNN,Faster RCNN and SG-Faster RCNN were outperformed by FSG-FD by over 4.5%accuracy at detecting occlusions.