Algorithms that Mimic the Brain's Processing Networks Preferred for Some Functions of Face Detection and Recognition Technology

• NCJ Number: 254467
• Date Published: March 2020
• Agencies: NIJ
• Grant Number(s): 2018-MU-CX-K111

Annotation

This article presents the methodology and findings of recently completed research funded by the National Institute of Justice (NIJ) that compares various types of facial analysis algorithms to improve technology used in face detection and recognition.

Abstract

Face detection and face recognition algorithms have improved in the past few years, largely due to the coupling of algorithms modeled on human brain processing functions called neural networks. This has facilitated increased computing power that provides rapid comparison of a viewed image with a dataset of millions of existing images. Still, there is doubt in the law enforcement and security communities regarding which algorithms work best for different face-analysis functions (e.g., face detection compared to face recognition) and in different face-comparison environments (e.g., small mugshot compared to fuzzy facial images or off-center shots from airport security cameras). More precise knowledge is needed on how alternative face algorithms work, so as to improve product performance and inform users and prospective buyers. Such an effort would also bring more fact-based evidence to the public discussion on the appropriate limits of face-recognition technology. The current project involved a rigorous comparison of face-analysis algorithms, including the Carnegie Mellon University (CMU) deep neural network algorithm, Ultron, which is designed to address difficult face-recognition challenges. The project concluded that the neural networks generally outperformed the conventional face-analysis algorithm of PittPatt and that CMU’s periocular reconstruction algorithm had better matching results than pre-existing products, warranting further research. More work is needed to better understand, measure, and compare competing algorithms. This will require more and better access to the “black box” of proprietary commercial algorithms, as well as a commitment to the performance of exacting, “apples-to-apples” comparisons of the strengths, weaknesses, and comparative merits of algorithms.