DEVELOPMENT OF LOW RESOLUTION FACE RECOGNITION SYSTEM WITH DEEP CONVOLUTIONAL NETWORK FOR REALTIME SURVEILLANCE

• Main Author: Andhika, Nixon
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55940
• Institution: Institut Teknologi Bandung

Face recognition system is a system that is capable of identifying an individual’s face based on the features of the face. Face recognition system have significantly improved in recent years and used for many applications, especially for surveillance. Face recognition system need high speed and good accuracy to be reliable enough to be used for surveillance. However, the performance of current state-of-the-art systems drops significantly when low-resolution face images is used. This, in turn, limits the application of face recognition system for surveillance that commonly only have low-resolution images. There are currently several low resolution face recognition systems, but they have several problems and their accuracies are still low. To solve these issues, this work develops a low resolution face recognition system with deep convolutional network. The system comprises of three components, which are face detection and extraction component, preprocessing component, and face recognition component. Experiments are done to determine the best techniques with good quality result and high speed for each component of the system. After comparison and evaluation at each component’s experiment, the best technique is chosen for the component. For face detection and extraction component, face detection technique SSD is chosen. For preprocessing component, super resolution technique CARN is chosen. For face recognition component, face embedding technique FaceNet is chosen with k-Nearest Neighbors classifier. The system is evaluated for recognition task using four versions of cleaned Labelled Faces in the Wild dataset, which are composed of one high-resolution version and three low-resolution versions with resolution of 32×32, 24×24, and 16×16 pixels. Based on the evaluation results, the developed system gains accuracy of 75.65% for 16×16 resolution, 96.08% for 24×24 resolution, 98.39% for 32×32 resolution, and 99.7% for high resolution. The system speed is 6 FPS for low-resolution images and 9 FPS for high-resolution images using Nvidia GTX 1050 GPU.