Abstract

Face mask-wearing identification systems automatically detect whether individuals wear masks in public spaces. However, existing CNN techniques face challenges in optimal hyperparameter selection, limiting their accuracy, interpretability, and computational efficiency. This study developed an Improved Pelican Optimization Algorithm integrated with Zebra Optimization Algorithm to optimize CNN hyperparameters (IPOA-CNN) for enhanced face mask identification. A dataset of 3,059 images from Kaggle was preprocessed using Contrast Limited Adaptive Histogram Equalization (CLAHE). The IPOA was developed by integrating ZOA’s exploration–exploitation mechanism and roulette wheel selection into the standard POA, then employed to optimize CNN parameters including weights, layer count, filter sizes, and number of filters. The framework was implemented in MATLAB R2023a and evaluated using accuracy, precision, sensitivity, specificity, false positive rate, and identification time. IPOA-CNN achieved 99.13% accuracy, 99.24% precision, 99.24% sensitivity, 99.24% specificity, 1.00% false positive rate, and 37.10 seconds identification time, outperforming POA-CNN (96.54%, 96.95%, 96.95%, 96.00%, 4.00%, 43.06s) and conventional CNN (94.08%, 94.67%, 94.91%, 93.00%, 7.00%, 65.76s). The IPOA-CNN model significantly outperforms conventional approaches by effectively optimizing critical CNN hyperparameters, making it suitable for real-time face mask recognition applications.