STARLIGHT Advances in AI-Powered Logo Detection for Criminal Investigations

STARLIGHT Advances in AI-Powered Logo Detection for Criminal Investigations

Dr. Henri Bouma from partner TNO introduced a STARLIGHT collaborative paper at the SPIE Conference on Artificial Intelligence for Security and Defence Applications in Amsterdam, held on September 4–5, 2023.

The research paper, titled "One-shot logo detection for large video datasets and live camera surveillance in criminal investigations," is authored by CERTH in collaboration with TNO and the Spanish National Police. It delves into significant advancements in the field of artificial intelligence for security and defense applications and introduces an innovative approach to logo detection with the potential to enhance forensic analysis.

Swiftly identifying key clues, such as logos on clothing, can be pivotal in speeding up criminal investigations, especially in video footage. Automatic logo detection expedites investigations, whether applied immediately after an incident on live camera streams or retroactively on extensive video datasets. Traditional methods required extensive training data—a vast dataset with numerous logo examples realistically annotated, a time-consuming and challenging task.

The STARLIGHT paper questions the conventional reliance on extensive training data by introducing an innovative approach demonstrating that effective logo detection can be achieved with just one or a few logo images used to train a deep neural network.

The approach consists of two key steps: data generation and logo detection. During data generation, a logo image seamlessly integrates into a person re-identification dataset, creating a synthetic dataset enriched with logos on clothing. The team employs various augmentation techniques to optimise performance. Subsequently, an object detector is trained on this synthetic dataset, enabling logo detections across various scenarios, including recorded images, video files, and live streams.

The innovation's significance is substantiated through quantitative and qualitative assessments. A comprehensive examination of the augmentation steps demonstrates the effectiveness of this approach. Additionally, qualitative evaluations from end users who tested the tool underscore its practicality and real-world impact.

This result marks a significant milestone in the integration of artificial intelligence and security. This innovative approach, which demands minimal training data, holds the potential to substantially enhance investigation speed and accuracy.

Please check out our Results section to explore more of the research from STARLIGHT.