DEVELOPMENT OF LOW-RESOLUTION FACE RECOGNITION SYSTEM USING SIAMESE NETWORK

• Main Author: Satyagama, Priagung
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51510
• Institution: Institut Teknologi Bandung

Low resolution face recognition system is a system that can recognize the identity of a face with input in the form of a low-resolution face image. The complete low-resolution face recognition system consists of a face detection component, a preprocessing or super resolution component, and a face recognition component. There are several low-resolution face recognition systems that have been developed before, but there are some disadvantages, such as incomplete system components and low recognition speed. The results of this study are intended to overcome these deficiencies. Therefore the purpose of this final project research is: 1) to build a low-resolution face recognition complete system that consists of at least face detection components, preprocessing components and face recognition components, 2) to compare quantitatively the performance of several deep learning techniques for each component built, 3) produce a low-resolution face recognition system using the most optimal components based on experiments conducted previously. System development begins with building a face detection component. The best face detection technique based on experiments on several face techniques is Mobilenet SSD, but in the system experiment the MTCNN technique will also be used as the second-best face detection technique to compare. Then the preprocessing or super resolution component is built. Based on experiments conducted, CARN was chosen as the best super resolution technique. Then a facial recognition component is built consisting of feature extraction sub-components and classifier. Each feature extraction technique to be compared is connected with a binary classifier. Based on experiments conducted on this component, Facenet was chosen as the best feature extraction technique. The overall system evaluation provides the best facial recognition accuracy performance of 93% in 36x36 images, 84% in 24x24 images, and 56.2% in 12x12 images. Whereas the best execution time is 0.084 seconds in 36 × 36 images, 0.086 seconds in 24 × 24 images, and 0.092 seconds in 12 × 12 images.