QCFace: Image Quality Control for boosting Face Representation & Recognition

TL;DR

QCFace introduces a hard-margin loss that explicitly encodes image recognizability into face embedding magnitudes while decoupling recognizability from identity. By coupling a Regularization Loss with a Recognizability Guidance term and a theoretical framework of margin strategies, it avoids mutual gradient interference that plagues soft-margin approaches and yields robust recognizability representations. The approach achieves state-of-the-art performance on verification and identification benchmarks, particularly on high- and mixed-quality datasets, and demonstrates improved distribution of feature magnitudes across a wide range. Limitations arise for very low-quality faces without large backbones or warm-up, but the method offers a principled path toward more reliable and interpretable face representations with practical impact on real-world FR systems.

Abstract

Recognizability, a key perceptual factor in human face processing, strongly affects the performance of face recognition (FR) systems in both verification and identification tasks. Effectively using recognizability to enhance feature representation remains challenging. In deep FR, the loss function plays a crucial role in shaping how features are embedded. However, current methods have two main drawbacks: (i) recognizability is only partially captured through soft margin constraints, resulting in weaker quality representation and lower discrimination, especially for low-quality or ambiguous faces; (ii) mutual overlapping gradients between feature direction and magnitude introduce undesirable interactions during optimization, causing instability and confusion in hypersphere planning, which may result in poor generalization, and entangled representations where recognizability and identity are not cleanly separated. To address these issues, we introduce a hard margin strategy - Quality Control Face (QCFace), which overcomes the mutual overlapping gradient problem and enables the clear decoupling of recognizability from identity representation. Based on this strategy, a novel hard-margin-based loss function employs a guidance factor for hypersphere planning, simultaneously optimizing for recognition ability and explicit recognizability representation. Extensive experiments confirm that QCFace not only provides robust and quantifiable recognizability encoding but also achieves state-of-the-art performance in both verification and identification benchmarks compared to existing recognizability-based losses.

Figures (11)

• Figure 1: Illustration of three types of hypersphere planning solutions for face representation learning. Constant margin shows the most lenient constraint in hypersphere planning, causing the least informative feature magnitudes. Although soft margin partially exploits the magnitude value, it possesses the entangled representation of recognizability. Hard margin (ours) facilitates positive correlation between magnitude value and confidence score to completely mitigate confusion in recognizability representation.
• Figure 2: Illustration of the difference in geometric interpretation between margin strategies in hypersphere planning. MVP is an abbreviation of magnitude value-driven planning. Circle and X points express well-planned and unexpected-planned embedded features. In (c), the blue feature has a large angular distance from its proxy, emphasizing ${\bm{g}}_\theta$ over ${\bm{g}}_{\|{\bm{z}}\|^-}$ and yielding $\|{\bm{g}}_{\|{\bm{z}}\|^+}\| \gg \|{\bm{g}}_{\|{\bm{z}}\|^-}\|$. Conversely, the black feature shows the opposite trend due to its smaller angular distance. In (d), we apply a unified gradient (i.e., ${\bm{g}}_{\|{\bm{z}}\|}$) to influence the magnitude of the embedded features, thereby disregarding the issue of unregulated gradients typically encountered in the soft margin approach. In (d), the feature's outline color indicates its expected region, while the center color represents its current region.
• Figure 3: The non-normalized geometrical representation of the feature space optimized by MagFace, AdaFace and QCFace-Arc. The feature vectors of each image extracted by face backbones are projected to the hyperplane formed by two proxies of two identities in the CASIA-WebFace dataset by the Gram-Schmidt algorithm. The borderred and green lines express the proxies of two identities. The actual proxy of each mapping feature point has the same color as that of feature. The gray dashed lines are the internal and external bisectors. The number below each image sample is the magnitude of the mapping feature vector.
• Figure 4: The magnitude distribution comparison between MagFace and QCFace-Arc.
• Figure 5: The training pipeline of face recognition models with class center loss.