Data-Specific Adaptive Threshold for Face Recognition and Authentication

Abstract

Many face recognition systems boost the performance using deep learning models, but only a few researches go into the mechanisms for dealing with online registration. Although we can obtain discriminative facial features through the state-of-the-art deep model training, how to decide the best threshold for practical use remains a challenge. We develop a technique of adaptive threshold mechanism to improve the recognition accuracy. We also design a face recognition system along with the registering procedure to handle online registration. Furthermore, we introduce a new evaluation protocol to better evaluate the performance of an algorithm for real-world scenarios. Under our proposed protocol, our method can achieve a 22% accuracy improvement on the LFW dataset.

Analyzing the effectiveness of image augmentations for face recognition from limited data

“…As the face recognition approach, we chose the FaceNet model with adaptive threshold. It shows near to state-of-the-art performance on our reference LFW dataset [17]. We used the official implementations of FaceNet 4 and Adaptive threshold 5 to construct our experiments.…”

“…It reduces the distance between an anchor image and a positive example of the same identity and increases the distance between the anchor and a negative example with a different identity. The current state-of-theart method is also based on the CNN [17] and has two operations in the system: registration and recognition. In the operation of registration, an embedding is extracted from an input face image by using a FaceNet model.…”

“…The most intuitive efficiency metric is accuracy, which is in general the number of correct observations to all observations. We chose a slightly modified accuracy metric that was introduced in the work of Chou, Lee, Chan, et al [17] and will be described below. An experiment starts with a test dataset of size N that consists of images I = {I 1 , I 2 , ..., I N } and labels P = {P 1 , P 2 , ..., P N }.…”

“…Small companies in the IT industry often lack access to huge corpora of data. For example, the current stateof-art FR technique FaceNet with adaptive threshold [17] has high accuracy results. However, it was trained on MS-Celeb-1M dataset [18] that is quite large.…”

Efficient Region of Interest Based Metric Learning for Effective Open World Deep Face Recognition Applications

“…The adaptive threshold schemes are recently used in different granularities like global threshold granularity, which determines a new threshold for the entire population. Another approach is to find a threshold for each category, such as gender and ethnicities within the population [31], or find a separate threshold for each person in FR application (subject or class) as proposed by chou et al [32].…”