RI: Small: Toward Human-Level Face Verification Performance Using Distinctive Features

点击翻译按钮获取中文摘要

This project aims to transform the way in which researchers approach face recognition technology by modeling distinctive features. Humans are capable of recognizing images of familiar faces even as they become extremely distorted. This project seeks to address current issues with face recognition technology by modeling the process after human perception. This project advances the fields of automated face recognition and human face perception by combining research in both areas to produce computational models of human face memory. This research will benefit society by producing techniques capable of recognizing faces in low-quality imagery as seen in surveillance and human-computer interaction settings. This project supports education and diversity through the recruitment of a diverse research team, the incorporation of research results into artificial intelligence courses and the wide dissemination of research results, data and code. This project is jointly funded by the Robust Intelligence (RI), the Established Program to Stimulate Competitive Research (EPSCoR), and the Secure and Trustworthy Cyberspace (SaTC) programs. This research investigates whether automated face verification performance can be improved by recognizing and emphasizing distinctive facial features. The project focuses on three main objectives: 1) modeling distinctive facial features, 2) face verification using distinctive features and 3) modeling exaggerated distinctive features. In modeling distinctive facial features, a new set of data will be collected with many images per identity and each identity labeled with distinctive features. Baseline and robust approaches to distinctive feature recognition will be developed and made publicly available along with the data. For face verification using distinctive features, multi-task learning approaches will be explored and evaluated on several large-scale surveillance and human-computer interaction datasets. The approach for modeling exaggerated distinctive features of faces involves learning generative models from weakly labeled data to produce realistic facial images from veridical faces. Automatically generated images will then be used to break up end-to-end deep learning frameworks for face verification in low-quality imagery.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.