After the COVID-19 pandemic, the effectiveness of face mask usage in such an environment has been seen. This situation triggered much research and experiments on this subject. Most of this research was done by using artificial intelligence. This paper focuses on a filter-based approach using a convolutional neural network for classification of face mask usage, which are mask correct, mask wrong, and no mask. Seventeen chosen filters are applied to the input images as an image pre-processing phase which results in 17 input images per image. Our customized network takes these 17 images as input and eliminates them into a single image before flattening and feeding to the dense layers. For training, 5 learning rates were tried starting from \(10^{-7}\) and increasing 10 times for each try stopping at \(10^{-3}\) . Results showed that increasing the learning rate shortened the training time and increased the accuracy. The highest test accuracy was obtained at \(10^{-3}\) learning rate with 98.86%. At \(10^{-4}\) learning rate, the second good result of 98.16% was obtained.