This study presents an Enhanced Mountain Gazelle Optimization driven Convolutional Neural Network (EMGO-CNN) for robust face based gender classification. The aim of the research is to improve classification reliability and generalization under varying decision thresholds by integrating an enhanced metaheuristic optimization strategy with deep feature learning. The primary objectives are to optimize CNN hyperparameters, reduce false classification rates, and achieve high sensitivity, specificity, and accuracy without increasing computational complexity. The proposed model was implemented and evaluated using MATLAB, providing a controlled and reproducible experimental environment for algorithm development and performance analysis. The dataset employed in this study consists of facial images from 5,330 individuals, including 2,480 male and 2,850 female samples, ensuring balanced representation and sufficient variability for effective model training and testing. The EMGO algorithm was used to guide the optimization of CNN parameters, enabling efficient exploration and exploitation of the search space. Performance evaluation was conducted across multiple decision thresholds to assess model stability and robustness under increasingly strict classification conditions. Experimental results demonstrate that at higher thresholds ranging from 0.36 to 0.98, the EMGO-CNN consistently maintained stable and highly competitive performance. At the optimal threshold of 0.51, the model achieved 2,457 true positives, 23 false negatives, 20 false positives, and 2,830 true negatives. These results correspond to the lowest observed false positive rate of 0.70% and the highest specificity of 99.30%. Sensitivity remained above 99.07%, while precision and overall accuracy reached 99.19%, confirming the model’s exceptional reliability even under strict decision cutoffs. Computational time showed minimal variation, ranging from 97.80 seconds at threshold 0.36 to 98.93 seconds at 0.51, indicating that performance improvements were achieved without significant efficiency loss. The results show that EMGO-CNN consistently outperformed the standard MGO-CNN by achieving superior sensitivity, specificity, and accuracy alongside a reduced false positive rate. These findings validate the effectiveness of the proposed framework as an optimized and reliable deep learning solution for face based gender classification tasks. The proposed approach demonstrates strong potential for deployment in real world biometric and intelligent vision systems across diverse applications.