PARTIAL FACE IDENTIFICATION ON A LIMITED POSE TYPE AND SAMPLE FACE DATASET

• Main Author: Djamaluddin, Muhammad
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79984
• Institution: Institut Teknologi Bandung

Since the early 1970s, face recognition has been a well-known research topic in computer vision, and it has advanced rapidly in the deep learning era. Face recognition has two primary tasks: face verification and face identification. Face recognition remains difficult due to four fundamental issues: pose, occlusion, facial expressions, and illumination. And three out of four of these real difficulties are part of the partial face problem, which makes facial recognition harder. One form of partial face is a profile face with a counterclockwise angle of 90, while a standard face perpendicular to the front of the camera is called a frontal face. Frontal face recognition is a closed problem that can be used as a reference for individual identity. The primary objective of this study is to investigate the difficulty of identifying a person's frontal face from his profile face when trained and tested on a database containing a very small number of facial images per subject and pose. For this research, the ITB Frontal Profile Limited Dataset (IFPLD) was created, which contained only one frontal face and one profile face for each subject. Many institutions still preserve individual data along with the individual's frontal face and profile faces in such personal databases. This study focused on addressing the challenge of face identification using a dataset with few samples. Two distinct methodologies were employed: a supervised learning approach and a meta-learning technique. The supervised learning strategy involves the extraction of features using two distinct methods: the classical method, namely Scale-Invariant Feature Transform (SIFT), and the deep learning method, specifically FaceNet. The input to the face verification system consists of concatenated FaceNet embedding vectors of frontal faces and profile faces. However, there is a significant imbalance in the proportions of data between the same face (minority) and various faces (majority). The classification system is developed using standard training methods and a method we call 2k-balanced training to ensure that the percentage of identical and different face class data is balanced. The experiments concluded that the face verification system has high accuracy but low precision and sensitivity. However, the supervised learning iv approach for face verification has low accuracy when used for face identification tasks, which are the main research topic. Utilizing a meta-learning technique, the second strategy looks at face identification in the IFPLD dataset as a few-shot learning problem, or more specifically, N-way- 1-shot learning, which is further studied with the help of Siamese and prototypical networks. The quality of the embedding vector features in both metric-based meta- learning techniques mostly determines the effectiveness of the final model. Fine- tuning was done on the ResNet-18 model, which has been pre-trained with ImageNet using a variety of training methods to obtain representative embedding vectors, such as using the siamese network with contrastive loss and triplet loss and SimCLR training method, which is a self-supervised learning (SSL) method, in order to produce representative features on the IFPLD dataset.. Two approaches to determining the distance between query data and prototypes are suggested in the prototypical network method, and they impact the accuracy level. By transforming 1-shot learning problems into k-shot learning, where there is k-1 augmentation data in training and testing episodes to gain superior performance, the prototypical network outperforms the siamese network in accuracy. It has been demonstrated that using augmentation data in prototype networks enhances model performance. The augmentation data comes from five basic image processing procedures. On the IFPLD dataset, the ResNet-18_IN* model produced the best results with 97%, 89%, and 82% accuracy in identifying 5, 20, and 40 individuals, respectively, in the N-way-1-shot learning schema.