Lightweight and accuracy upgrade: Surface partial Key point detection model based on deep reinforcement learning

Abstract

This paper introduces a novel lightweight and high-accuracy model for facial key point detection, incorporating advancements in deep learning to optimize performance under limited computational resources. The proposed architecture, Inception-MLP, integrates a pre-trained Inception module for multi-scale feature extraction and replaces traditional fully connected layers with a single-hidden-layer multilayer perceptron (MLP), reducing complexity while maintaining strong representational power. Unlike conventional CNN-based models that often suffer from parameter redundancy and prolonged training times, this method strikes an effective balance among accuracy, model size, and training efficiency. Comparative evaluations against Basic CNN, VGG, and EfficientNet-Lite3 demonstrate that Inception-MLP offers superior performance across multiple metrics, making it suitable for real-world applications, especially on mobile or embedded devices. The model's ability to achieve precise localization of facial landmarks with fewer parameters highlights its deployment potential in scenarios requiring both lightweight inference and high reliability. This work contributes a feasible direction toward efficient deep learning-based facial analysis, providing a valuable reference for future model design in similar tasks.