Researchers have introduced a face recognition training method that combines two measures of image quality with separate mechanisms for pulling images of the same person together and pushing different identities apart.
El Ouanas Belabbaci of Algeria’s University of Sciences and Technology Houari Boumediene and Bhavesh Wani and Philipp Terhörst of Johannes Gutenberg University Mainz presented Dual Quality Margin Learning for Face Recognition, or DQM-Face, at the European Conference on Computer Vision on September 12.
Modern face recognition models often treat the magnitude of a facial feature vector as an estimate of image quality. The researchers argue that magnitude alone can be distorted by blur, occlusion or lighting and does not capture whether an image contains identity-specific information.
DQM-Face adds a semantic quality branch using squeeze-and-excitation attention, a technique that learns which feature channels carry useful information. It combines that semantic score with the magnitude score to adjust the attraction margin used during training. A second, curriculum-based margin increases separation between different identities.
The team trained the system on the MS1MV2 face dataset with an iResNet-100 backbone and evaluated it on image and video benchmarks. DQM-Face achieved an average verification accuracy of 97.58 percent across LFW, AgeDB-30, CFP-FP and CPLFW, including 98.83 percent on the cross-pose CFP-FP test.
The authors said the system ranked first or second across their principal comparisons on those datasets and the larger IJB-B and IJB-C benchmarks. They also cautioned that differences among leading methods on the smaller image benchmarks are too narrow to be statistically significant, making the larger evaluations a more robust test.
Benchmark design has become a central part of biometric research. An IJCB competition measured gains from synthetic training data, while the FBI has used identical twins to probe the limits of facial recognition under unusually difficult matching conditions.
The semantic branch adds about 99,000 parameters to a 65.2-million-parameter backbone and less than 0.0001 GFLOPs of computation, according to the paper. Quality estimation uses features already produced by the recognition network, and the researchers have released the implementation as open-source code.
ECCV placed the work in its open-world recognition, segmentation and transfer track.
Sources: ECCV 2026, DQM-Face paper
–
By the ID Tech Editorial Team
