Model Capacity for Small-Sample Facial Expression Recognition

Abstract

Facial expression recognition (FER) is increasingly required in classroom affect analysis, lightweight human-computer interaction, and domain-specific behavioral monitoring, where only limited labeled facial images are available. Reliable FER in such low-resource settings is critical because unstable predictions can distort downstream decisions, particularly for minority classes and visually similar negative expressions. Existing pipelines often increase depth, attention, region fusion, or transfer learning capacity, but the completed experiments show that these designs can mismatch small datasets and underperform compact alternatives. This paper presents a stability-oriented convolutional study that rethinks model capacity for seven-class small-sample FER. The framework combines stratified partitioning, grayscale normalization, compact VGG-style representation learning, global average pooling, label smoothing, dropout, L2 regularization, and momentum-based optimization to control estimator variance while preserving local expression cues. An error-aware protocol further evaluates macro-F1, weighted-F1, class-wise recall, sensitivity behavior, and confusion patterns. Experiments on a 981-image FER dataset with a stratified 70/15/15 split compare LeNet, VGG variants, ResNet-small, ROI-fusion VGG, and MobileNetV2 transfer learning. Results show that VGG-optimized achieves 97.30% accuracy and 95.64% macro-F1, while lightweight LeNet reaches 98.65% accuracy, and complex ROI/transfer alternatives drop to 56.08% and 83.78%, confirming that capacity control dominates architectural complexity.