Abstract
This paper presents an optimized version of the MobileNetV3 architecture tailored for real-time employee emotion recognition. By integrating deep separable convolutions, inverted residual structures, and a novel serial spatial-channel attention mechanism, the model achieves high accuracy while maintaining computational efficiency. Experimental results on the KDEF dataset demonstrate a peak accuracy of 94.95%, outperforming standard lightweight models and several state-of-the-art architectures.
Introduction: The Need for Efficient Emotion Recognition
Monitoring employee well-being is essential for organizational health, yet traditional manual observation is subjective and unscalable. While facial expression recognition (FER) via deep learning offers a solution, high-performing models often require significant computational resources, making them unsuitable for deployment on edge devices like office tablets or smart cameras. This research addresses the gap between high-level accuracy and the need for lightweight, real-time processing in workplace environments.
Methodology: Serial Spatial-Channel Attention
The core innovation of this study is the modification of the MobileNetV3 framework. While the original model utilizes Squeeze-and-Excitation (SE) blocks to focus on channel-wise information, the authors propose a Serial Spatial-Channel Attention (Serial-SA-CA) module.
As shown in the architecture above, the module first processes spatial attention to identify where in the image the most relevant emotional features are located. It then sequentially processes channel attention to determine which feature maps are most significant. This serial approach allows for a more refined extraction of nuanced facial features compared to parallel configurations. Additionally, the network's time-consuming layers were optimized to reduce power consumption and memory usage without compromising feature extraction capabilities.
Experimental Setup and Datasets
The model was evaluated using the Karolinska Directed Emotional Faces (KDEF) dataset, which contains approximately 4,900 images of 70 individuals. The dataset covers seven fundamental emotional states: happiness, sadness, disgust, anger, surprise, fear, and neutrality. KDEF is particularly valuable for benchmarking due to its standardized lighting and specific facial instructions, which minimize environmental noise during training.
Training utilized a 7:3 train-test split with Adam and SGD optimizers. Transfer learning was employed to leverage pre-trained weights, accelerating convergence and improving generalization across different facial types.
Key Results and Discussion
The improved MobileNetV3 (Serial-SA-CA) demonstrated superior performance across multiple metrics. Notably, the serial configuration of the attention module was found to be more effective than a parallel one, achieving 94.95% accuracy compared to 93.31% for the parallel version.
As illustrated in the table, the Serial-SA-CA-MobileNetV3 outperformed the standard SE-MobileNetV3 (94.81%) and even edged out the more complex POSTER++ model (94.44%). The model showed exceptional specificity (above 98%), effectively minimizing false positives—a critical requirement for workplace monitoring where misidentifying a neutral expression as a negative one could lead to unnecessary interventions.
While the model excelled at identifying "Neutral" (99.51%) and "Joy" (98.57%), it faced more challenges with "Fear" (89.25%) and "Sadness" (90.28%). This suggests that certain emotions share subtle facial muscle movements that remain difficult for lightweight architectures to distinguish perfectly.
Conclusion and Future Work
The proposed improvements to MobileNetV3 provide a robust, efficient solution for real-time emotion recognition in professional settings. However, the authors note that real-world deployment must address privacy and ethical concerns regarding sensitive data collection. Future research will focus on testing the model's robustness in "in-the-wild" office environments with varying lighting and occlusions, as well as exploring multimodal approaches that combine facial analysis with voice or physiological signals.
Summary updated