Generiranje sintetičnih prstnih sledi

Fingerprints collected from various surfaces, also known as fingermarks, are important evidence for identifying subjects that were present on a crime scene. With the recent advances in deep learning aproaches, methods have been developed to identify fignermarks and match them to subjects in police databases. However, training such a method has proven to be difficult due to a lack of training data. Such datasets are difficult and expensive to collect, and come with privacy concerns. In this work, we explore the suitability of diffusion models, which have recently gained popularity in image generation tasks, for the task of generating a dataset of varied and realistic fingermark impressions of a known identity. We finetune a latent diffusion model using low-rank adaptation and ControlNet guidance. We show that our approach is capable of generating high-quality and varied samples after being trained on only 20 images of fingermarks for each style. We also show that ControlNet can be used to guide the generation of fingermarks towards a wanted identity, despite it being trained only on synthetic images.