Lightweight Micro-Expression Recognition Based on Structured and Unstructured Pruning

Micro-expression refers to facial expressions that last for an extremely short time, typically less than 0.5 seconds. It can reveal a person's true emotion, which is one of the powerful indicators to judge what a person actually thinks. However, the deployment of micro-expression recognition algorithms onto resource-constrained devices poses a challenge due to their high computational requirements. In this essay, we optimize the identification of microexpression algorithms by using pruning-based lightweight methods. By using unstructured pruning and structured pruning methods, the backbone Resnet-18 network model is subjected to local and global lightweight pruning, respectively, where the network structure is optimized to reduce the time and space costs required for micro-expression recognition algorithms. The experimental results demonstrate that the original algorithm's performance is enhanced following the pruning procedure with fewer parameters, and it exhibits superior performance on the CASMEII, SMIC, and SAMM datasets.