AI based criminal identification system using facial recognition

This paper presents the development of a criminal identification system that leverages on deep learning for face recognition to enhance accuracy, speed and reliability in law enforcement. The system was developed using the Object-Oriented Analysis and Design Methodology (OOADM) to guide the requirements modeling, system structuring, and implementation. At its core, the system employed a Residual Neural Network (ResNet) architecture, chosen for its ability to extract deep hierarchical features and maintain robust recognition under challenging conditions such as poor lighting, pose variations, and partial occlusions. A diverse dataset of facial images was used for training and evaluation, with preprocessing steps including face detection, resizing, normalization, and alignment to ensure consistency. The model size using Feature Descriptor Parameters was optimized, and sensitive data were tokenized using Universally Unique Identifier (UUID). Additionally, the National Identification Number (NIN) dataset was simulated. To validate performance, this deep learning-based pipeline Haar + ResNet model was compared with the classical pipeline, that is, the traditional Haar Cascade face detector combined with Local Binary Patterns Histograms (LBPH). While Haar + LBPH performed adequately under controlled conditions, it showed reduced accuracy with difficult lighting and non-frontal images. In contrast, ResNet achieved consistently higher accuracy and faster matching times, making it better suited for real-world applications. Beyond recognition, the system cross-referenced the identified individuals with the NIN registry official records for identity verification. Experimental results confirmed the model’s high precision and reliability in supporting its potential deployment in surveillance and investigative systems.