TeLL Me what you can't see - a forensic analysis for images data augmentation

Face recognition systems have become indispensable in modern criminal investigations, enabling law enforcement agencies to rapidly identify and track suspects. As technological advancements grow rapidly, face recognition systems are expected to play an even more crucial role in enhancing public safety and criminal justice processes. This growth is hindered by the quality of images at disposal of security agencies. Poor quality of mugshots causes face recognition systems to often encounter issues with false positives, highlighting the need for stricter regulations and technological improvements to ensure the fair and effective use of mugshots in modern law enforcement. This thesis wants to propose a forensic tool that can be used for data augmentation in the process of a criminal investigation. Starting from low quality images, the tool aims to generate a description of the subject represented in the image and to generate a new synthetic picture maintaining an high level of ID fidelity to the original subject. The purpose of the tool proposed is to automate and speed up the process of generation of criminal wanted posters and also reduce biases in criminal investigations correlated to low quality images, that can actively influence human witnesses. To do so, different deep neural networks and machine learning technique are investigated. Specifically in the context of images quality enhancement, multimodal reasoning and computer vision. This research aims to show the process undergone to the creation of the tool, going from the selection of a dataset suited to the task, to the definition of methodologies for the evaluation of results. As an addition to what afore mentioned, it is also explored the aging process of a subject represented in an image. This is done through the definition of an independent dataset, the research of the tool’s networks different applications and an extensive evaluation of results. Experimental results show that the proposed tool proves to be effective in achieving an image data augmentation in the majority of cases leading to high success probabilities.