This paper describes an integrated approach to fatigue classification that combines a face detector based on the multi-task convolutional neural network (MTCNN) algorithm with a binary image classification model. An MTCNN algorithm that extracts facial regions from input images is developed, pre-processes these images, and augments them. A convolutional neural network composed of convolutional and fully connected layers will analyze facial features to discriminate between “active” and “fatigue” states. The architecture proposed demonstrates robustness through data augmentation and dropout regularization. An Adam optimizer and a binary cross-entropy loss, designed in the training strategy, have been utilized. In the performance evaluation, all the above-mentioned assessment metrics are used, including training accuracy curves, confusion matrices, receiver operator characteristics analysis, and precision-recall. The effectiveness of the architecture is validated on a dataset of active and fatigue states, showing its capability for real-world applications.