Erasing tattoos from photos via AI could aid facial recognition systems
Researchers from the Darmstadt University of Applied Sciences in Germany have recently published a new paper detailing a method to use artificial intelligence to remove tattoos from photos of people, with potential biometrics implications.
According to lead researcher Mathias Ibsen, face painting and tattoos can potentially impair the performance of computer vision systems like facial recognition algorithms by altering what models “expect to see” in key areas of the face.
To better understand and mitigate the effect of facial tattoos in facial analysis systems, Ibsen and his team believe an increased number of large datasets of images of individuals with and without tattoos are needed.
To this end, the researchers propose a generator for automatically adding realistic tattoos to facial images.
“To create a facial image with tattoos, the face is first divided into face regions using landmark detection whereafter tattoo placements can be found,” the paper reads.
“Subsequently, deep reconstruction maps and cut-out maps can be estimated from the input image. Thereafter, the information is combined to realistically blend tattoos to the facial image.”
They then demonstrated the feasibility of the generation by training two deep learning-based models for removing tattoos from face images.
The models used were pix2pix, a supervised conditional generative adversarial network (GAN) for image-to-image translation, and SkinDeep, a pre-existing model for removing body tattoos utilizing well-known components which have shown “good results for other image-to-image translation tasks.”
Results using the novel mode reportedly showed that it was possible to remove facial tattoos from real images without degrading the quality of the image.
Moreover, the paper suggested that, by using the proposed deep learning-based tattoo removal before extracting and comparing facial features, facial recognition accuracy was increased overall.
“The comparison scores are not significantly affected for the pix2pix and SkinDeep models which only showed moderate capabilities of removing tattoos from facial images,” the paper reads.
“However, for SkinDeep, which has been trained on the synthetic database, it is shown that the dissimilarity score on average gets lower, which indicates that the recognition performance might improve.”
The technique could presumably also be applied to overcome attempts to defeat facial recognition systems with makeup.