Lightweight and High-Accuracy Facial Landmark Detection Network by Applying Artificial Intelligence

Facial emotion recognition (FER) for tracking and authentication has become a hot topic due to the rapid development of AI and its widespread application. Many different scenarios could benefit from portable facial recognition technology. However, due to limits in size, weight, power source, and notably computing capability, current lightweight mobile phones struggle to meet the accuracy and real-time needs of face recognition monitoring. Thus, issues like face identification monitoring and lightweight implementation are crucial and need fixing quickly. In response to this issue, this paper presents Tiny-HRN+, a lightweight facial recognition method. By plummeting the quantity of parameters and the volume of processing needed, the model is made easier to construct with the help of FN's partial convolution technique. In order to satisfy the demand for high-precision recognition, it employs a parameter-free attention strategy that boosts the model's recognition accuracy by increasing its complexity. We have added a dynamic prediction head to the model to improve its predictive skills, allowing it to better fulfill the needs of detection performance. This has been empirically proven using the freely accessible Wider-Face (WF) dataset. The proposed face recognition method in this paper demonstrates clear lightweight benefits, including the provided variable and calculation loads, the model size of 9.1 MB, an accuracy of 91.15%, and an FPS of 68, and visual verification of the outcomes of detection reveals a noticeably better efficiency compared to other models.