This study developed an enhanced mountain gazelle optimized convolutional neural network for a face-based gender recognition system. The specific objectives formulate an Enhanced Mountain Gazelle Optimizer Technique (EMGO) using Chaotic Exponential Map function, and design an optimized Convolutional Neural Network Technique for face-based gender recognition using the formulated EMGO. The acquisition of MP4 and AVI video datasets were obtained as primary data from YouTube. It comprises 5,330 face samples extracted from the YouTube, including 2,480 male and 2,850 female faces, with each video contributing between 554 and 823 detected faces per gender category. Face detection was carried out using the Viola–Jones algorithm, followed by preprocessing operations these are resizing, cropping, grayscale conversion, and adjustment of brightness and contrast to enhance image quality. The result shows the false positive rate (FPR) of the three models CNN, MGO-CNN, and EMGO-CNN across different threshold values. It is evident that the baseline CNN records the highest FPR values, starting at 4.00% at threshold 0.2 and only reducing slightly to 3.75% at threshold 0.75. Also, it was shown the specificity performance of CNN, MGO-CNN, and EMGO-CNN at various threshold values. The baseline CNN maintains a stable but comparatively lower specificity, ranging from 96.00% at a threshold of 0.2 to 96.25% at a threshold of 0.75. This indicates that CNN is less capable of correctly identifying true negatives, meaning it tends to misclassify some negative samples as positive. In conclusion, the developed EMGO-CNN model has demonstrated superior performance compared to traditional CNN and MGO-CNN approaches for face-based gender recognition system. By integrating enhancements into the Mountain Gazelle Optimization algorithm, the EMGO framework improved exploration and exploitation capabilities, thereby preventing premature convergence and enabling a more reliable selection of optimal CNN hyperparameters. Therefore, the EMGO-CNN model is recommended for real-world deployment in gender recognition systems where accuracy, speed, and robustness are critical.