3D FACE VERIFICATION FROM SINGLE LOW RESOLUTION IMAGE BASED ON EXPLOITATION DEEP CONVOLUTIONAL FEATURE

• Main Author: Qomalita Hijriana, Zalid
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/59466
• Institution: Institut Teknologi Bandung

A biometric system uses a unique part of the human body to verify, for example, the face. Face is naturally the most commonly used for verification, but face is also a very complex part due to many internal and external factors that affect facial features, such as lighting, camera position, facial expressions, and the presence of occlusion. Moreover, the use of surveillance cameras in public areas nowadays is increasing. Facial images obtained from surveillance cameras usually are in a low resolution because they are taken from a long distance. Low-resolution images are difficult to verify because they lose a lot of important information from the image. In this study, the super-resolution method was applied to deal with the problem of low-resolution input images. Evaluation of three super-resolution methods, namely Efficient Sub-Pixel Convolutional Neural Network (ESPCN), Enhanced Deep Super-Resolution (EDSR), and Fast Super-Resolution Convolutional Neural Network (FSRCNN), was conducted to choose the best super-resolution method for face verification task. To reduce the effect of pose variations, 3D features were added. It was obtained from the reconstruction of 2D images into 3D and then 3D feature extraction was carried out with Deep Convolutional Neural Network (DCNN) with the addition of Squeeze and Excitation Network (SE) blocks. The results showed that verification using the FSRCNN super-resolution method followed by 3D reconstruction and 3D feature extraction based on the DCNN+blockSE configuration, with network depth 26 and stride 2, obtain 80.07% in accuracy. This result is 11.71% higher than the baseline using FaceNet with bicubic interpolation.