DCM-Net: A deep composite micro-tuning network for enhancing feature generalization in low-resolution facial expression recognition

College of Innovation Engineering, Macau University of Science and Technology, Macau, 999078, China

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Driven by technological advancements, computers have become capable of recognizing and classifying human facial expressions in images and further inferring the emotional state of individuals. This technology, known as facial expression recognition, has now become one of the most challenging and attractive research directions in the field of computer vision. This paper addresses the core issue of “insufficient feature generalization ability in facial expression recognition of low-resolution images” and proposes an innovative solution - Deep Composite Micro-tuning Network (DCM-Net). This solution includes a triple optimization design: backbone architecture optimization, feature enhancement design, and classification boundary construction. The VGG16 convolutional group is adopted to extract high-level semantic features; Two-stage convolutional refinement layers are added to extract expression-sensitive representations, while introducing a dense regularization chain to enhance feature expression; The overfitting problem is solved through an average pooling strategy, and a hierarchical fully connected layer is utilized to construct the classification boundary. Comprehensive comparative experiments were conducted on three neural network architectures, including the implementation of the core of stochastically connected units and the implementation of the stochastic weight averaging algorithm. The results show that DCM-Net outperforms the comparative networks in both recognition accuracy and computational efficiency.