Efficient GRU-based Facial Expression Recognition with Adaptive Loss Selection

Abstract

As real-world deployment of facial expression recognition systems becomes increasingly prevalent, computational efficiency emerges as a critical consideration alongside recognition accuracy. Current research demonstrates pronounced emphasis on accuracy maximization through sophisticated convolutional architectures, yet systematic evaluation of efficiency-performance trade-offs remains insufficient for practical deployment scenarios. This paper addresses this gap through comprehensive analysis of recurrent neural network architectures for facial expression recognition, specifically comparing Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) implementations within a novel one-vs-all classification framework incorporating adaptive loss function selection. A rigorous 2 × 2 × 2 factorial experimental design systematically evaluates architecture (GRU vs LSTM), optimization strategy (Bayesian vs predefined), and loss function complexity (standard vs advanced with auto-selection) across six basic emotions using the CK+ dataset with MediaPipebased facial landmark features. The investigation reveals that GRU architectures achieve statistical performance equivalence with LSTM while demonstrating 25% computational efficiency advantage (relative complexity 0.75 vs 1.0). The proposed adaptive loss selection mechanism automatically selects focal loss for severe class imbalance (ratio > 11.5), weighted binary cross-entropy for moderate imbalance (ratio 3.5-11.5), and standard binary cross-entropy otherwise. System performance achieves 92.7% ± 5.0% overall accuracy, with per-emotion F1-scores exhibiting substantial variability from 0.215 (fear) to 0.967 (surprise). Comprehensive statistical analysis incorporating power analysis and practical equivalence testing demonstrates optimization strategy equivalence across 25% of evaluated metrics, while architectural comparisons reveal non-equivalence despite similar performance levels. The study acknowledges significant limitations including critically small sample size (n=6 per condition), single dataset validation, and theoretical rather than empirical efficiency validation. These findings provide evidence-based guidelines for architecture selection in resource-constrained facial expression recognition applications, with the adaptive loss selection framework representing a significant methodological contribution for addressing class imbalance challenges in emotion recognition systems.