AI-KD: Towards Alignment Invariant Face Image Quality Assessment Using Knowledge Distillation

TL;DR

This work tackles the sensitivity of face image quality assessment (FIQA) to alignment by introducing Alignment-Invariant Knowledge Distillation (AI-KD). AI-KD uses a teacher–student knowledge-distillation setup where a frozen teacher backbone $M^t$ guides a trainable student backbone $M^s$ and an $MLP$ quality regressor to predict quality scores, while on-the-fly alignment perturbations generate misaligned samples $\,hat{x}_i$ from properly aligned images $x_i$. The training optimizes both a representation loss $\,\mathcal{L}_{cos}$ between $e_i = M^t(x_i)$ and $\,\hat{e}_i = M^s(\hat{x}_i)$ and a quality loss $\,\mathcal{L}_1$ between $q_i$ and $\,\hat{q}_i$, with perturbations defined as $\,hat{k}^{pts}_i = k^{pts}_i + \mathcal{U}_{[-p,p]}$ and $p=3$. Across six datasets, four FR models, and seven FIQA baselines, AI-KD yields consistent improvements, extending robustness to misaligned inputs and often achieving state-of-the-art results for properly aligned data as well. The approach thereby reduces dependence on a specific landmark detector and enhances FIQA utility in real-world, unconstrained capture scenarios. The method’s practicality is underscored by extensive experiments and public code release.

Abstract

Face Image Quality Assessment (FIQA) techniques have seen steady improvements over recent years, but their performance still deteriorates if the input face samples are not properly aligned. This alignment sensitivity comes from the fact that most FIQA techniques are trained or designed using a specific face alignment procedure. If the alignment technique changes, the performance of most existing FIQA techniques quickly becomes suboptimal. To address this problem, we present in this paper a novel knowledge distillation approach, termed AI-KD that can extend on any existing FIQA technique, improving its robustness to alignment variations and, in turn, performance with different alignment procedures. To validate the proposed distillation approach, we conduct comprehensive experiments on 6 face datasets with 4 recent face recognition models and in comparison to 7 state-of-the-art FIQA techniques. Our results show that AI-KD consistently improves performance of the initial FIQA techniques not only with misaligned samples, but also with properly aligned facial images. Furthermore, it leads to a new state-of-the-art, when used with a competitive initial FIQA approach. The code for AI-KD is made publicly available from: https://github.com/LSIbabnikz/AI-KD.

Figures (1)

• Figure 1: Overview of the proposed Alignment Invariant Knowledge Distillation (AI-KD) process. The proposed approach trains a quality-regression model, consisting of a FR backbone $M^s$ and a quality regression head $MLP$, on quality labels $q_i$ extracted using any existing FIQA approach. Training samples $\hat{x}_i$ are (mis)aligned on the fly, by perturbing the correct landmark $k^{pts}_i$ of the initial unaligned face samples $\ddot{x}_i$. Additionally, to ensure robustness to alignment variations in the distilled model, we design a distillation objective that ensures consistency between representations of the aligned $e_i$ and (mis)aligned images $\hat{e}_i$, as well as matching the predicted quality scores $\hat{q}_i$ to the quality labels $q_i$.