Face Image Quality Vector Assessment for Biometrics Applications

In this paper, we propose a multi-task convolutional neural network which produces an image quality vector for an input face image. This image quality vector contains the face quality score and the information about the nuisance factors (i.e., pose, illumination, blurriness and expression) that caused the predicted image quality score. Our multitask network utilizes a pretrained ResNet-50 as its stem. Employing different data augmentation techniques, we create a huge and diverse dataset. We ground truth this dataset and use it to fine-tune our multi-task neural network. Our Multi-task learning framework enables us to learn a shared and beneficial feature representation among the relevant tasks to achieve a better performance. Moreover, the proposed multi-task neural network provides useful information about the nuisance factors. Nuisance factors information is useful for applications like face image quality assessment during the automatic enrollment process where user's photo should comply with a standardized criteria. In this case, if the user upload a low-quality image which does not comply with a predefined standard, the system provides feedback about the nuisance factors associated with the captured image. Therefore, the user can resolve the problem, and upload a new image that can remedy the issue. Although an extensive research has been done on face image quality assessment, non of them have addressed face image quality vector effectively. To the best of our knowledge, our method is the first method that uses deep learning approach to generate a face image quality vector. The evaluation of results demonstrates that our method gets higher or comparable accuracy for face image quality assessment in comparison to the state-of-the-art methods and analyzes the input image to provide detailed information about the nuisance factors.