State-of-the-art architectures of convolutional neural networks (CNN) are widely used by authors for facial expression recognition (FER). There are many variants of these models with positive results in studies for FER and successful applications, some well-known models are VGG, ResNet, Xception, EfficientNet, DenseNet. However, these models have considerable complexity for some real-world applications with limitations of computational resources. This paper proposes a lightweight CNN model based on a modern architecture of dense-connectivity with moderate complexity but still ensures quality and efficiency for facial expression recognition. Then, it is designed to be integrated into learning management systems (LMS) for recording and evaluation of online learning activities. The proposed model is to run experiments on some popular datasets for testing and evaluation, the results show that the model is effective and can be used in practice.