LITE-FER: A Lightweight Facial Expression Recognition Framework for Children in Resource-Limited Devices

This study proposes a lightweight facial expression recognition (FER) framework for children that can be used on resource-limited devices such as socially assistive robots interacting with children in real world applications. In this study, knowledge distillation (KD) and unstructured weight pruning (UWP) method are used to achieve lightweight FER models. Effect of joint usage of KD and UWP method on FER is also evaluated by using a pruned teacher model within the teacher-student training paradigm in knowledge distillation method. Experiments are performed utilizing AffectNet, CK+ and CAFE datasets. LITE-FER achieved 89.69% and 77% in CK+ and CAFE respectively in k-fold cross validation strategy. LITE-FER only consists of 113.98K parameters (445.24 KB) which makes the model resource-efficient. LITE-FER can reach up to 173.82 FPS with TensorRT on GPU, and 30.09 FPS with keras on CPU. LITE-FER revealed its maximum inference performance in batch inference with 3213 FPS throughput. Results showed that our proposed model (“LITE-FER”) results in comparable accuracy, as well as computational and memory efficiency. The proposed model is planned to be used in assistive robotic systems for children in the future.