GraFIQs: Face Image Quality Assessment Using Gradient Magnitudes

TL;DR

This work tackles the need for effective face image quality assessment without labeled quality data or regression training. It introduces GraFIQs, a training-free FIQA method that measures how much BN statistics differ between training and a test sample, then backpropagates this BN-discrepancy to generate gradient magnitudes whose sum serves as the quality score. By using the gradient signal from a pretrained FR model, GraFIQs achieves competitive performance across diverse benchmarks and FR models, often surpassing IQA baselines and matching SOTA FIQA methods while requiring only a single forward and backward pass. This gradient-magnitude based, BN-discrepancy approach provides a practical, label-free avenue for FIQA with broad applicability to real-world FR systems.

Abstract

Face Image Quality Assessment (FIQA) estimates the utility of face images for automated face recognition (FR) systems. We propose in this work a novel approach to assess the quality of face images based on inspecting the required changes in the pre-trained FR model weights to minimize differences between testing samples and the distribution of the FR training dataset. To achieve that, we propose quantifying the discrepancy in Batch Normalization statistics (BNS), including mean and variance, between those recorded during FR training and those obtained by processing testing samples through the pretrained FR model. We then generate gradient magnitudes of pretrained FR weights by backpropagating the BNS through the pretrained model. The cumulative absolute sum of these gradient magnitudes serves as the FIQ for our approach. Through comprehensive experimentation, we demonstrate the effectiveness of our training-free and quality labeling-free approach, achieving competitive performance to recent state-of-theart FIQA approaches without relying on quality labeling, the need to train regression networks, specialized architectures, or designing and optimizing specific loss functions.

Figures (24)

• Figure 1: An overview of the proposed GraFIQs for assessing the quality of unseen testing samples. Sample $\mathcal{I}$ is passed into the pretrained FR model and BNS are extracted. Then, the MSE between the BNS obtained by processing the testing sample and the one recorded during the FR training is calculated. The MSE is backpropagated through the pretrained FR to extract the gradient magnitudes of parameter group $\phi$. Finally, the absolute sum of gradient magnitudes of $\phi$ is calculated and utilized as FIQ.
• Figure 2: EDC for FNMR@FMR=$1e-3$ and FNMR@FMR=$1e-4$ of our proposed method using $\mathcal{L}_{\text{BNS}}$ as backpropagation loss and absolute sum as FIQ. The gradients at image level ($\phi=\mathcal{I}$), and block levels ($\phi=\text{B}1$$-$$\phi=\text{B}4$) are used to calculate FIQ. $\text{MSE}_{\text{BNS}}$ as FIQ is shown in black. Results are shown on benchmarks Adience, AgeDB30 and XQLFW datasets using ArcFace model. The proposed GraFIQs method leads to lower verification error when images with the lowest utility score estimated from gradient magnitudes are rejected. Furthermore, estimating FIQ by backpropagating $\mathcal{L}_{\text{BNS}}$ yields significantly better results than using $\text{MSE}_{\text{BNS}}$ directly.
• Figure 3: EDC curves for FNMR@FMR=$1e-3$ for all evaluated benchmarks using MagFace, ElasticFace, and CurricularFace FR models. AUC are shown in Table \ref{['tab:erc_sota_comparison']}. EDC curves for ArcFace are provided in the supplementary material. The proposed GraFIQs method, shown in solid red, utilizes gradient magnitudes and it is reported using the best setting from Table \ref{['tbl:ablation']}.
• Figure 4: Error-versus-Discard Characteristic (EDC) curves for FNMR@FMR=$1e-3$ and FNMR@FMR=$1e-4$ of our proposed method using $\mathcal{L}_{\text{BNS}}$ as backpropagation loss and absolute sum as FIQ. The gradients at image level ($\phi=\mathcal{I}$), and block levels ($\phi=\text{B}1$$-$$\phi=\text{B}4$) are used to calculate FIQ. $\text{MSE}_{\text{BNS}}$ as FIQ is shown in black. Results shown on benchmark Adience Adience using ArcFace, ElasticFace, MagFace, and, CurricularFace FR models. It is evident that the proposed GraFIQs method leads to lower verification error when images with lowest utility score estimated from gradient magnitudes are rejected. Furthermore, estimating FIQ by backpropagating $\mathcal{L}_{\text{BNS}}$ yields significantly better results than using $\text{MSE}_{\text{BNS}}$ directly.
• Figure 5: Error-versus-Discard Characteristic (EDC) curves for FNMR@FMR=$1e-3$ and FNMR@FMR=$1e-4$ of our proposed method using $\mathcal{L}_{\text{BNS}}$ as backpropagation loss and absolute sum as FIQ. The gradients at image level ($\phi=\mathcal{I}$), and block levels ($\phi=\text{B}1$$-$$\phi=\text{B}4$) are used to calculate FIQ. $\text{MSE}_{\text{BNS}}$ as FIQ is shown in black. Results shown on benchmark AgeDB30 agedb using ArcFace, ElasticFace, MagFace, and, CurricularFace FR models. It is evident that the proposed GraFIQs method leads to lower verification error when images with lowest utility score estimated from gradient magnitudes are rejected. Furthermore, estimating FIQ by backpropagating $\mathcal{L}_{\text{BNS}}$ yields significantly better results than using $\text{MSE}_{\text{BNS}}$ directly.