Generating synthetic tattoo dataset for biometric applications

Tattoos have proven to be valuable tools for law enforcement agencies in identifying both criminals and victims. Although the availability of datasets containing images of tattooed individuals is limited due to ethical and privacy issues, such datasets could be useful in forensics, for instance, to train neural networks that can identify and classify tattoos automatically. A potential solution to this limitation is the generation of synthetic datasets using virtual reality environments (i.e., Unity). In our work, we proposed an approach for developing datasets containing potentially infinite images of both tattooed and non-tattooed synthetic realistic subjects. We also introduced a method to automatically generate segmentation masks to localize the tattoos within the images. To enhance realism, the resulting images were further improved photorealistically using commercial generative AI. The efficacy of the datasets was evaluated by implementing convolutional neural networks to discriminate between tattooed and non-tattooed subjects. Experimental results show that this approach can be used for image classification, demonstrating that synthetic images can serve as a starting point for generating image classifiers, while more photorealistic images result in lower classification errors. The codes for generating the dataset in the VR environment and the generated images have been made available at https://github.com/IreneDeZotti/VRTattooDataset .