18 Mar 2021
09:00 Doctoral defense Fully distance
Theme
Face Recognition Exploring Part-Based Representations
Student
Marcus de Assis Angeloni
Advisor / Teacher
Helio Pedrini
Brief summary
In recent years, there has been an increasing use of biometrics in our daily lives, such as unlocking cell phones, electronic commerce and bank authentication. Face recognition has several advantages over other biometric modalities, since it is natural, non-intrusive and is a task that people perform routinely and effortlessly. Despite significant recent advances in face recognition, there are still open challenges, such as in situations with occlusion in part of the face, different poses, blurs in the image, facial expressions, different lighting and the use of makeup. These situations are commonplace in surveillance applications and images captured by mobile devices. As demonstrated by the cognitive science literature, there is a high probability that part-based processing may exist in the perception of the face by humans. However, despite this evidence, the facial recognition literature still focuses on holistic processing and representations derived from global face image alignments. In this thesis, we investigate the adoption of facial parts in the task of biometric face recognition inspired by this observation, proposing different strategies to cut out these parts, choose appropriate representations for each one and how to combine their results. We approach the adoption of facial parts as an incremental work, whose methodology and idea have been refined, since facial recognition is well researched and evolves quickly. First, we explored traditional representations of facial parts (eyebrows, eyes, nose and mouth) to cover controlled and uncontrolled scenarios, using four public databases. The results of the fusion experiments show that specific features and classifiers for each facial part can improve the accuracy of biometric systems. Second, we propose a compact multi-input convolutional neural network (CNN) architecture to explore end-to-end learning from facial parts to age estimation from a single image captured in a real world setting. Experiments conducted on a public database show that our method has competitive precision with a reduced number of parameters and input size. Finally, we focus on the challenging problem of facial verification with makeup. We propose two strategies for cutting facial parts (around fiducial points or facial thirds), extracting representations using state-of-the-art CNN models and merging their scores with the holistic score. Experiments carried out in four databases and also in a cross-protocol show that improvements have been achieved in the evaluation metrics, even without retraining or fine-tuning the CNN models. Once we have made available the source code of the proposed approaches, the experiments can be reproduced to regenerate and extend the results obtained. In addition, the results obtained indicate some new directions and research topics in facial recognition.
Examination Board
Headlines:
| Hélio Pedrini | IC / UNICAMP |
| Erickson Rangel do Nascimento | DCC / UFMG |
| Ronaldo Cristiano Prati | CMCC / UFABC |
| Esther Luna Colombini | IC / UNICAMP |
| Alexandre Mello Ferreira | IC / UNICAMP |
Substitutes:
| Gerberth Adín Ramírez Rivera | IC / UNICAMP |
| João Paulo Papa | FC / UNESP |
| Moacir Antonelli Ponti | ICMC / USP |