تحديد هوية الشخص بناءً على تحليل الطب الشرعي لشيخوخة الوجه باستخدام شبكات عصبية عميقة

Person Identification Based On Facial Aging Forensics Analysis Using Deep Neural Networks

Digital Image Forensics (DIF) is an area of knowledge focused on recovering and analyzing digital evidence in a criminal investigation process based on employing a system for computer vision. The recent advancement of technology, biological properties have become a necessity for person identification and verification systems. Forensic person identification is an important field that examines the variety of biological indicators that are available to investigators. The facial identification system has some challenges that worsen the performance of identification such as facial expression variation, illumination variation, pose variation, occlusion, and the hard challenge of facial aging. Face identification with aging is an extremely fascinating and difficult topic of pattern recognition research. This thesis focus on person identification with aging challenge. In general, the aging process significantly alters the shape and skin texture of the face, making it extremely harder to properly identify a person. Face identification across age is therefore a current open issue in the field of computer vision. It has a wide range of exciting real-time applications, such as the capacity to renew passports and licenses, find children who have gone missing or persons who have Alzheimer’s, identify criminals, and more.
The proposed work presents a person identification system based on facial aging forensics analysis using deep neural networks. The main aim of the proposed system is to predict the person’s identity by his/her face at different ages. The proposed system is built by concatenate two models representing the model-specific function stage performed respectively (face aging model and person identification model) each one contained several steps that performed different functions.
The first one based on CycleGAN model is the face aging that is responsible for coming up with a representation and matching scheme for face that is robust to changes due to facial aging based on extracting salient features that can achieve temporal invariance. This model provides the identification of younger face from older one, as well as has the ability to do the inverted operation, i.e. reverting from young to old by using and training CycleGAN to do forward aging and reverse aging.
The second model based on Mobilenet V2 is the identification that determines the identity of the person based on extracting facial essential features that can clearly show the individual characteristics that decide the person’s identity. The modify pre-trained Mobilenet V2 is used to build the second model in order to gain accurate results by only modifying some layers in the original architecture to identity the identity of 105 people. Each model is preceded by a number of preprocessing steps.
These preprocessing steps includes: image resize (scale) and data normalization as well as the Region of Interest ROI detection and extraction before face aging model.
Because there is no dataset that satisfied the purpose of the proposed system, we are used two sample datasets during proposed system stages. These datasets are: UTKFace used with face aging model and 105-pins-classes samples dataset with person identification model. The enrichment is done by merging between samples of these two datasets and the accuracy attained 98%.