Lightweight facial expression recognition based on joint model compression

In this paper, we propose a lightweight facial expression recognition algorithm based on joint model compression. Firstly, we propose a self-distillation method based on facial expression recognition, which not only considers the overall output of the model but also focuses on the prediction results of each facial expression category, thereby improving the performance and generalization ability of the facial expression recognition model. Secondly, in order to further compress the facial expression recognition model, we apply a channel pruning method based on transformed <i>L</i>₁ regularization. This pruning method imposes non-convex constraints on the parameters of the facial expression recognition model through transformations of the parameters in the <i>L</i>₁ norm penalty term, achieving more accurate sparse regularization and effectively preventing overfitting of the model. Lastly, the pruned network is quantified to low-bit to obtain the final lightweight facial expression recognition model. The proposed method achieves 73.25&percnt; on the FER2013 dataset, with a parameter compression rate of approximately 50&percnt;, which achieves an effective lightweight.