We introduce a novel Face Image Quality Assessment (FIQA) methodology that focuses on the stability of face embeddings in the embedding space, evaluated through controlled perturbations. Our approach is predicated on the hypothesis that the quality of facial images is directly correlated with the stability of their embeddings; lower-quality images tend to exhibit greater variability upon perturbation. We have evaluated our model on three different FR models and three different datasets. We have also compared the results to state-of-the-art methods.
This schematic depicts the process of assessing image quality through a novel perturbation technique. Initially, the original image undergoes controlled perturbations to create a variant, symbolizing potential quality degradation. Both the original and perturbed images are then transformed into the embedding space using a Face Recognition (FR) model. In this space, we compute the cosine similarity between the embeddings of the original and perturbed images. The core hypothesis of our technique is that for lower-quality images, perturbations in the input space lead to more significant deviations in the embedding space, thereby resulting in a lower similarity score between the original and perturbed images. This stability of image embedding forms the backbone of our FIQA methodology. The quality scores displayed on the image are obtained using affine perturbation with the ArcFace FR model on the XQLFW database.
In our experimentation, we systematically explored a variety of perturbation techniques, including affine transformations, Poisson noise, Gaussian noise, and salt-and-pepper noise. Additionally, we investigated the effects of both random and structured occlusions, as well as elastic transformations and mixup transformations.
In our study, we compared various perturbation-based FIQA techniques with top state-of-the-art methods using non-interpolated EDC curves. Our findings reveal that affine transformation, in particular, shows remarkable performance, closely rivaling leading methodologies like DifFIQA and FaceQAN. These results underscore the potential of perturbation approaches in enhancing FIQA systems and set new benchmarks for future research in the field.
For detailed information, see our seminar paper: seminar paper.
This project uses weights for the ArcFace, AdaFace, and TransFace models. The weights for these models must be downloaded and placed in their respective directories as described below:
- ArcFace weights: Place in
face_recognition/arcface. Refer to the ArcFace paper for more details: ArcFace: Additive Angular Margin Loss for Deep Face Recognition. - AdaFace weights: Place in
face_recognition/adaface. Refer to the AdaFace paper for more details: AdaFace: Quality Adaptive Margin for Face Recognition. - TransFace weights: Place in
face_recognition/transface. Refer to the TransFace paper for more details: TransFace: Translating a Face Embedding to a Recognizable Image.
After adding the weights, ensure that the paths to these weight files are correctly set in the utils and main files. This is crucial for the proper functioning of the face recognition models.
To run the experiments, navigate to the /experimental_work folder and execute the main.py script.