Panasonic Connect Co. Ltd, Panasonic R&D Center Singapore (Singapore Research Institute), and Singapore’s Nanyang Technological University (NTU) have developed a new facial recognition technology that has been proven effective in improving the accuracy of facial recognition with non-Caucasians and females, which has been a conventional and industry-wide issue due to the smaller data sets that have traditionally been available for training facial recognition models.

A research paper on the new method has been accepted for publication at the International Conference on Computer Vision (ICCV) 2023, the top conference in the field of image recognition, highlighting the international value of the new technology.

Conventional facial recognition technology tends to have low recognition accuracy for non-Caucasian races and women due to smaller data sets used to train the recognition models. Lower recognition accuracy rates can thus lead to errors or even false identifications. To address this problem, the three organizations have developed a technology named “Invariant Feature Regularization for Fair Face Recognition” that applies the ‘partition-based invariant learning‘ method of deep learning to facial recognition for the first time. By learning a facial recognition model that is commonly effective for various attributes, this technology has succeeded in reducing the error rate for races with small data sets and the error rate for women, without compromising the accuracy for races with large data sets.

Panasonic Connect is committed to creating fair environments for everyone, regardless of nationality or gender. The new technology reflects our continuous effort to ensure that facial recognition applications can be utilized equally and fairly, regardless of race or gender.

Research background

With the facial recognition technology provided by Panasonic Connect, the company aims to create fair conditions for all. However, conventionally a general industry-wide problem has been that the accuracy of facial recognition tends to be lower for races and genders for which there are small data sets to train from. The training data for facial recognition reflects real-world population proportions, so the scale of the data varies between races and between genders.

Deep learning technology, which forms the basis of facial recognition technology, requires large amounts of training data and is affected by differences in the size of this data, resulting in differences in accuracy between attributes (races and genders).

Proposed method

To alleviate the effects of this bias in facial learning data, the joint development focused on a technique in deep learning called partition-based invariant learning. The method automatically divides the facial training data into groups (partitions) for each attribute, as shown in the diagram below, using the difficulty of recognition as an indicator. It also learns invariant feature representation that improves the accuracy of all partitions.

By performing this partition-based invariant learning multiple times, feature representation invariably available for various partitions, such as race and gender, is learnt — i.e., a facial recognition model that is effective for a variety of attributes is learnt step by step.

This method combined with the conventional facial recognition method achieved the highest accuracy with an average accuracy for four racial groups (African, Caucasian, South Asian and East Asian) in the Masked Face Recognition Challenge, an evaluation data set that can verify the accuracy for each racial group. In addition, it reduced the female acceptance error rate compared to the conventional face recognition method on CelebA, an assessment dataset that validates accuracy by gender.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved