Abstract
Deaths and injuries are common in road accidents, violence, and natural disasters. In such cases, one of the main tasks of responders is to retrieve the identity of the victims to reunite families and ensure proper identification of deceased or injured individuals. Apart from this, identification of unidentified dead bodies due to violence and accidents is crucial for police investigation. In the absence of identification cards, current practices for this task include DNA profiling and dental profiling. Face is one of the most commonly used and widely accepted biometric modalities for recognition. However, face recognition is challenging in the presence of facial injuries such as swelling, bruises, blood clots, laceration, and avulsion which affect the features used in recognition. In this paper, for the first time, we address the problem of injured face recognition and propose a novel Subclass Contrastive Loss (SCL) for this task. A novel database, termed as Injured Face (IF) database, is also created to instigate research in this direction. Experimental analysis shows that the proposed loss function surpasses existing algorithms for injured face recognition.
Introduction: The Challenge of Facial Trauma
Standard face recognition (FR) systems have reached high levels of accuracy on "clean" datasets. However, physical trauma—sustained in accidents, sports, or violence—distorts the facial geometry and texture that these models rely on. This creates a significant "domain gap" where existing models, trained on non-injured faces, fail to generalize to injured states.
In forensic and emergency scenarios, identifying unconscious or deceased individuals is critical. Currently, responders often rely on DNA or dental profiling, which are slow and costly. Automated FR could provide a faster alternative if it can be made robust to the appearance changes caused by injury. As shown in the figure above, popular pre-trained models like VGGFace and OpenFace perform very poorly on injured faces, with Rank-1 accuracies often falling below 10%.
Methodology
The authors propose that an injury should not be treated as random noise, but as a systematic transformation. They introduce the Subclass Contrastive Loss (SCL), which treats "injured" and "non-injured" states as distinct subclasses of a single identity.
Subclass Contrastive Loss (SCL)
The SCL framework operates through two primary mechanisms:
- Intra-class Minimization: The loss function minimizes the distance between the non-injured subclass and the injured subclass of the same subject. It also reduces variance between different types of injured images for that individual.
- Inter-class Maximization: It maximizes the distance between the subclasses of different subjects to ensure discriminative feature learning.
The system uses a LightCNN-9 architecture in a Siamese-style network, where weights are shared between branches to learn a robust, injury-invariant embedding space.
The Injured Face (IF) Database
To support this research, the authors curated the Injured Face (IF) database, containing 898 images of 100 subjects (346 injured, 552 non-injured). Images were sourced from the internet, covering a wide range of trauma types from sports, accidents, and violence.
Results and Discussion
Experimental results demonstrate that the proposed SCL significantly improves performance over standard training objectives on the IF database:
- Performance Gains: SCL achieved a Rank-1 Accuracy of 36.70%, outperforming standard Contrastive Loss (28.67%) and Triplet Loss (30.26%).
- Baseline Comparison: The best-performing pre-trained model (LightCNN-9) reached only 25.43% Rank-1 accuracy before being fine-tuned with the proposed SCL.
- Scalability: In an "Extended Gallery" test adding 100 additional distractor identities from the SCface dataset, the model maintained a stable accuracy of 36.13%, suggesting robustness to larger search spaces.
Qualitative analysis shows that SCL effectively reduces the overlap between "genuine" and "imposter" score distributions, making the system more reliable for verification tasks.
Conclusion
This work represents the first dedicated effort to recognize faces under severe physical trauma. By explicitly modeling injuries as subclasses, the SCL framework learns features that remain invariant even under significant distortion. While a ~37% accuracy indicates that injured face recognition remains an arduous task, this research provides a foundational dataset and methodology for future improvements in forensic identification.
Summary updated