Optimizing MobileNetV3 for multimodal eye gaze and emotion recognition via advanced pruning and quantisation techniques

This study introduces an innovative approach for optimising visual understanding by leveraging MobileNet V3 for eye-gazing, eye-blinks, and emotional expressions recognition. The importance of visual cues, such as gaze direction, blinking patterns, and facial expressions, in various applications, including human-computer interaction and behavioural analysis, is driving the need for efficient and accurate models that operate effectively in resource-constrained environments. MobileNet V3 offers a strong foundation for such tasks due to its lightweight architecture; however, it can be enhanced further to provide even greater performance. To achieve this, we apply advanced model optimisation techniques, including pruning and quantisation, to reduce computational complexity without compromising accuracy. We validate our approach using three distinct datasets: EyeGaze, Emotions, and Closed Eye, which offer diverse visual inputs across different scenarios. The results demonstrate that our optimised MobileNet V3 model accurately detects and analyses eye gaze, blinks, and emotional expressions, making it a robust algorithm for real-world applications. All the codes for reprehensibility and trained models can be found at our github repository .