Robust thermal face recognition for law enforcement using optimized deep features with new rough sets-based optimizer

Abstract

In the security domain, the growing need for reliable authentication methods highlights the importance of thermal face recognition for enhancing law enforcement surveillance and safety especially in IoT applications. Challenges like computational resources and alterations in facial appearance, e.g., plastic surgery could affect face recognition systems. This study presents a novel, robust thermal face recognition model tailored for law enforcement, leveraging thermal signatures from facial blood vessels using a new CNN architecture (Max and Average Pooling- MAP-CNN). This architecture addresses expression, illumination, and surgical invariance, providing a robust feature set critical for precise recognition in law enforcement and border control. Additionally, the model employs the NM-PSO algorithm, integrating neighborhood multi-granulation rough set (NMGRS) with particle swarm optimization (PSO), which efficiently handles both categorical and numerical data from multi-granulation perspectives, leading to a 57% reduction in feature dimensions while maintaining high classification accuracy outperforming ten contemporary models on the Charlotte-ThermalFace dataset by about 10% across key metrics. Rigorous statistical tests confirm NM-PSO’s superiority, and further robustness testing of the face recognition model against image ambiguity and missing data demonstrated its consistent performance, enhancing its suitability for security-sensitive environments with 99% classification accuracy.