Mutual Information Maximization for Facial Expression Recognition in Images
Wenrui Dou, Kunxia Wang · 2023
In human daily communication, only 7% of information is transmitted through language, while 55% of information is conveyed through facial expressions. Nowadays, the increasing number of only children, the serious aging of the population and other problems, the topic of empty-nest elderly has also attracted much attention. Therefore, applying FER research to focus on the psychological state of the elderly is also very important. With age, facial attributes such as shape and surface texture also undergo certain changes over time. To achieve good performance in elderly facial recognition tasks, a key point depends largely on how to effectively use representations to capture facial features. In order to achieve this goal, we use the Vision Transformer model, which uses the Vit model to obtain features in shallow and deep layers, and introduces MI limits to enhance their relationship with image content. We impose the constraint of local mutual information maximization (LMIM) to require that the features generated at each time step have a strong relationship with the image content. We evaluated it on three large benchmarks. Finally, we conducted a Specific analysis and comparison of the model. Our results proved the validity of the proposed method, the experimental results after incorporating LMIM improving by about 2% to 9% compared to the original results.