Enhanced Micro-Expression Recognition Using MBCC-CNN Architecture

Abstract

This study presents an innovative approach for micro-expression recognition based on a multi-branch cross-connected convolutional neural network (MBCC-CNN). Unlike conventional single-stream architectures, the proposed framework effectively captures and enhances image features through a multi-branch design integrated with residual connections, Network-in-Network modules, and hierarchical (tree-based) structures. A cross-linked shortcut mechanism is incorporated to merge outputs from different convolutional layers, facilitating smoother inter-branch information flow, expanding the receptive field, and strengthening feature extraction. This design enables efficient fusion of branch-level features, overcoming the limitations of inadequate learning in isolated branches and significantly improving recognition accuracy. Experimental evaluations conducted on benchmark datasets—SAMM, CASME I, CASME II, and CAS(ME)²—achieved recognition rates of 99.89%, 98.73%, 100.00%, and 99.58%, respectively. Comparative analysis with state-of-the-art approaches demonstrates that the proposed MBCC-CNN architecture achieves superior recognition performance and enhanced robustness in micro-expression analysis.