Nowadays, facial expression recognition (FER) has drawn considerable attention from the research community in various application domains due to the recent advancement of deep learning. In the education field, facial expression recognition has the potential to evaluate students' engagement in a classroom environment, especially for deaf and hard-of-hearing students. Several works have been conducted on detecting students' engagement from facial expressions using traditional machine learning or convolutional neural network (CNN) with only a few layers. However, measuring deaf and hard-of-hearing students' engagement is yet an unexplored area for experimental research. Therefore, we propose in this study a novel approach for detecting the engagement level ('highly engaged', 'nominally engaged', and 'not engaged') from the facial emotions of deaf and hard-of-hearing students using a deep CNN (DCNN) model and transfer learning (TL) technique. A pre-trained VGG-16 model is employed and fine-tuned on the Japanese female facial expression (JAFFE) dataset and the Karolinska directed emotional faces (KDEF) dataset. Then, the performance of the proposed model is compared to seven different pre-trained DCNN models (VGG-19, Inception v3, DenseNet-121, DenseNet-169, MobileNet, ResNet-50, and Xception). On the 10-fold cross-validation case, the best-achieved test accuracies with VGG-16 are 98% and 99% on JAFFE and KDEF datasets, respectively. According to the obtained results, the proposed approach outperformed other state-of-the-art methods.