A face photo forms a fundamental element of almost every identity document such as national ID cards, passports, etc. The governmental agencies issuing such documents may set slightly different requirements for a face image to be acceptable. Nevertheless, some are too critical to avoid, such as mouth closedness, eyes openness and no veil-over-face. In this paper, we aim to address the problem of fully automating the inspection of these 3 characteristics, thereby enabling the face capturing devices to determine, as soon as a face image is taken, if any of them is invalid or not. To accomplish this, we propose a deep learning-based approach by defining model architectures that are lightweight enough to enable real-time inference on resource-constrained devices with a particular focus on prediction accuracy. Lastly, we showcase the performance and efficiency of our approach, which is found to surpass two well-known off-the-shelf solutions in terms of overall precision.