A Study on Facial Landmark Detection and Image Processing

This research focuses on developing a robust face alignment system crucial for applications in facial recognition, expression analysis, and augmented reality. Utilizing a dataset of 2811 training images and 554 test images each annotated with 44 facial landmarks, the study explores several models including Convolutional Neural Networks (CNN), Cascade Regression, and Residual Networks (ResNet). Through rigorous experimentation, it was observed that while CNNs and Cascade Regression models provided substantial accuracy, the ResNet model displayed superior performance, especially in complex scenarios involving diverse expressions and occlusions. Additionally, the project also developed a simple algorithm for lip and eye color modification, further broadening the applicability of the system in image processing. Results indicate that while traditional models like CNN and Cascade Regression perform adequately, ResNet offers superior accuracy and robustness, particularly in challenging conditions. This research confirms the effectiveness of ResNet in enhancing face alignment technologies and suggests potential areas for future improvements in the field.