This study presents a comparative evaluation of two Convolutional Neural Network (CNN) architectures, ResNet50 and MobileNetV3-Large, for multiclass face mask classification under various environmental conditions. Unlike most previous studies that primarily evaluate model performance on clean images, this research focuses on analyzing model robustness against controlled synthetic environmental perturbations using a balanced subset of the Kaggle Face Mask Dataset consisting of 9,156 images. Both models were implemented using a transfer learning approach and evaluated under normal conditions as well as five types of environmental variations, namely brightness enhancement, darkening, Gaussian blur, JPEG compression, and rotation, each applied at five severity levels with three repeated runs using different random seeds. Model performance was assessed using accuracy, precision, recall, and F1-score. The experimental results show that ResNet50 achieved a baseline accuracy of 94.32%, slightly outperforming MobileNetV3-Large with an accuracy of 94.10%. As the severity of environmental perturbations increased, both models experienced performance degradation, with darkening and Gaussian blur producing the largest reductions in classification accuracy, while JPEG compression and rotation had relatively minor effects. Overall, ResNet50 demonstrated greater robustness across most environmental variations, whereas MobileNetV3-Large offered superior computational efficiency through fewer parameters, a smaller model size, and faster inference time. These findings highlight the trade-off between robustness and computational efficiency in CNN-based face mask classification, suggesting that the choice of architecture should be guided by application requirements. Nevertheless, the findings are limited to a single public dataset and controlled synthetic perturbations, indicating the need for future studies using more diverse datasets and real-world image acquisition conditions.