Development of an Optimized Extreme Learning Machine Model for Face-Based Crime Control Recognition System Using Mantis Search Algorithm

Abstract

Face recognition can be used for crime control and public security, due to its ability to identify individuals without physical contact. However, variations in illumination, pose, image quality and occlusion affect its accuracy, worsened in real-world settings where surveillance cameras rarely capture a clear shot of a face. A fast, simple classifier, Extreme Learning Machine (ELM), was introduced, but it suffers from random initialization of its hidden-layer weights and biases, producing inconsistent accuracy. This led to proposing an optimized model, Mantis Search Algorithm-Extreme Learning Machine (MSA-ELM), which searches for the ELM configuration giving the best accuracy rather than choosing randomly. Images were acquired from the Federal School of Surveying, Oyo, comprising 300 individuals with three images each (900 images), grouped into crime-suspect and non-crime-suspect classes. The dataset was split 70% training and 30% testing, pre-processed by resizing, cropping and grayscale conversion, then features extracted using Local Binary Pattern (LBP) prior to classification. The model was implemented in MATLAB R2016a with an interactive GUI for training, testing and identification. MSA-ELM achieved 95.93% accuracy, 3.42% FAR and 4.58% FRR against 92.22%, 7.69% and 7.84% for traditional ELM, reducing recognition time from 73.79 to 55.87 seconds, supporting its consideration for real-world surveillance and suspect-identification deployment.