Unsupervised Face Recognition using Unlabeled Synthetic Data

Klemt, Marcel; Fang, Meiling; Kuijper, Arjan; Damer, Naser

doi:10.1109/fg57933.2023.10042627

Cited by 22 publications

(10 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The pretrained model produced a synthetic face data for each latent vector that was randomly non-repeatedly generated based on Gaussian noise. These images were then filtered automatically by using the CR-FIQA [3] face image utility assessment approach, where extreme non-frontal poses and largely occluded images were mostly removed by removing the images with the lowest utility score. This helps simulate the real log-in face recognition scenario that is commonly targeted by PAs.…”

Section: Synthaspoof Datasetmentioning

confidence: 99%

“…One of the main candidate solutions for this issue is the use of synthetic data [8]. This has been very recently and successfully proposed for the training of face recognition [4,5,30] and morphing attack detection [9,11,21], among other processes such as model quantization [2]. Synthetic data for PAD development has, besides the privacy and legal motivations, a major advantage when it comes to scale and diversity.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

SynthASpoof: Developing Face Presentation Attack Detection Based on Privacy-friendly Synthetic Data

Fang¹,

Huber²,

Damer³

2023

Preprint

View full text Add to dashboard Cite

Recently, significant progress has been made in face presentation attack detection (PAD), which aims to secure face recognition systems against presentation attacks, owing to the availability of several face PAD datasets. However, all available datasets are based on privacy and legallysensitive authentic biometric data with a limited number of subjects. To target these legal and technical challenges, this work presents the first synthetic-based face PAD dataset, named SynthASpoof, as a large-scale PAD development dataset. The bona fide samples in SynthASpoof are synthetically generated and the attack samples are collected by presenting such synthetic data to capture systems in a real attack scenario. The experimental results demonstrate the feasibility of using SynthASpoof for the development of face PAD. Moreover, we boost the performance of such a solution by incorporating the domain generalization tool MixStyle into the PAD solutions. Additionally, we showed the viability of using synthetic data as a supplement to enrich the diversity of limited authentic training data and consistently enhance PAD performances. The SynthASpoof dataset, containing 25,000 bona fide and 78,800 attack samples, the implementation, and the pre-trained weights are made publicly available 1 .

show abstract

Section: Synthaspoof Datasetmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

SynthASpoof: Developing Face Presentation Attack Detection Based on Privacy-friendly Synthetic Data

Fang¹,

Huber²,

Damer³

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…For all of the methods, higher quality score output is intended to indicate higher biometric utility [1]. Six of these were modern deep learning models specifically intended to assess quality in terms of image utility [1] for face recognition: CR-FIQA(S) & CR-FIQA(L) [15] (respectively with iResNet50/iResNet100 backbone trained on CASIA-WebFace [16]/MS1MV2 [7], both with 112 × 112 input image size), MagFace [17] (iResNet100 backbone trained on MS1MV2 [7], 112 × 112 input image size), SER-FIQ [18] ("same model" variant using ArcFace, 112 × 112 input image size), FaceQnet-v0 [19] & FaceQnet-v1 [20] (both with ResNet-50 backbone, trained on VGGFace2 [21], 224 × 224 input image size). These six models are all publicly available.…”

Section: Quality Assessmentmentioning

confidence: 99%

Evaluating Face Image Quality Score Fusion for Modern Deep Learning Models

Schlett

Rathgeb

Tapia

et al. 2022

2022 International Conference of the Biometrics Special Interest Group (BIOSIG)

View full text Add to dashboard Cite

Lossy face image compression can degrade the image quality and the utility for the purpose of face recognition. This work investigates the effect of lossy image compression on a state-of-the-art face recognition model, and on multiple face image quality assessment models. The analysis is conducted over a range of specific image target sizes. Four compression types are considered, namely JPEG, JPEG 2000, downscaled PNG, and notably the new JPEG XL format. Frontal color images from the ColorFERET database were used in a Region Of Interest (ROI) variant and a portrait variant. We primarily conclude that JPEG XL allows for superior mean and worst case face recognition performance especially at lower target sizes, below approximately 5kB for the ROI variant, while there appears to be no critical advantage among the compression types at higher target sizes. Quality assessments from modern models correlate well overall with the compression effect on face recognition performance.

show abstract

“…The system used a StyleGAN network, pretrained without labeled data to train a prototypical network that can identify human faces. Paper [ 5 ] uses GAN-based augmentation, which learns to maximize the similarity between two augmented images of the same synthetic instance, leading to high recognition accuracy. The unsupervised autonomous learning system, which learns from a video sequence, is presented in paper [ 6 ].…”

Section: Introductionmentioning

confidence: 99%

Autonomous Face Classification Online Self-Training System Using Pretrained ResNet50 and Multinomial Naïve Bayes

Maciura¹,

Cieplak

Pliszczuk³

et al. 2023

Sensors

View full text Add to dashboard Cite

This paper presents a novel, autonomous learning system working in real-time for face recognition. Multiple convolutional neural networks for face recognition tasks are available; however, these networks need training data and a relatively long training process as the training speed depends on hardware characteristics. Pretrained convolutional neural networks could be useful for encoding face images (after classifier layers are removed). This system uses a pretrained ResNet50 model to encode face images from a camera and the Multinomial Naïve Bayes for autonomous training in the real-time classification of persons. Faces of several persons visible in a camera are tracked using special cognitive tracking agents who deal with machine learning models. After a face in a new position of the frame appears (in a place where there was no face in the previous frames), the system checks if it is novel or not using a novelty detection algorithm based on an SVM classifier; if it is unknown, the system automatically starts training. As a result of the conducted experiments, one can conclude that good conditions provide assurance that the system can learn the faces of a new person who appears in the frame correctly. Based on our research, we can conclude that the critical element of this system working is the novelty detection algorithm. If false novelty detection works, the system can assign two or more different identities or classify a new person into one of the existing groups.

show abstract

Unsupervised Face Recognition using Unlabeled Synthetic Data

Cited by 22 publications

References 26 publications

SynthASpoof: Developing Face Presentation Attack Detection Based on Privacy-friendly Synthetic Data

SynthASpoof: Developing Face Presentation Attack Detection Based on Privacy-friendly Synthetic Data

Evaluating Face Image Quality Score Fusion for Modern Deep Learning Models

Autonomous Face Classification Online Self-Training System Using Pretrained ResNet50 and Multinomial Naïve Bayes

Contact Info

Product

Resources

About